tools/perf/util/header.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <errno.h>
   3 #include <inttypes.h>
   4 #include "util.h"
   5 #include "string2.h"
   6 #include <sys/param.h>
   7 #include <sys/types.h>
   8 #include <byteswap.h>
   9 #include <unistd.h>
  10 #include <stdio.h>
  11 #include <stdlib.h>
  12 #include <linux/compiler.h>
  13 #include <linux/list.h>
  14 #include <linux/kernel.h>
  15 #include <linux/bitops.h>
  16 #include <linux/stringify.h>
  17 #include <sys/stat.h>
  18 #include <sys/types.h>
  19 #include <sys/utsname.h>
  20 #include <unistd.h>
  21
  22 #include "evlist.h"
  23 #include "evsel.h"
  24 #include "header.h"
  25 #include "memswap.h"
  26 #include "../perf.h"
  27 #include "trace-event.h"
  28 #include "session.h"
  29 #include "symbol.h"
  30 #include "debug.h"
  31 #include "cpumap.h"
  32 #include "pmu.h"
  33 #include "vdso.h"
  34 #include "strbuf.h"
  35 #include "build-id.h"
  36 #include "data.h"
  37 #include <api/fs/fs.h>
  38 #include "asm/bug.h"
  39 #include "tool.h"
  40
  41 #include "sane_ctype.h"
  42
  43 /*
  44  * magic2 = "PERFILE2"
  45  * must be a numerical value to let the endianness
  46  * determine the memory layout. That way we are able
  47  * to detect endianness when reading the perf.data file
  48  * back.
  49  *
  50  * we check for legacy (PERFFILE) format.
  51  */
  52 static const char *__perf_magic1 = "PERFFILE";
  53 static const u64 __perf_magic2    = 0x32454c4946524550ULL;
  54 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
  55
  56 #define PERF_MAGIC      __perf_magic2
  57
  58 const char perf_version_string[] = PERF_VERSION;
  59
  60 struct perf_file_attr {
  61         struct perf_event_attr  attr;
  62         struct perf_file_section        ids;
  63 };
  64
  65 struct feat_fd {
  66         struct perf_header      *ph;
  67         int                     fd;
  68         void                    *buf;   /* Either buf != NULL or fd >= 0 */
  69         ssize_t                 offset;
  70         size_t                  size;
  71         struct perf_evsel       *events;
  72 };
  73
  74 void perf_header__set_feat(struct perf_header *header, int feat)
  75 {
  76         set_bit(feat, header->adds_features);
  77 }
  78
  79 void perf_header__clear_feat(struct perf_header *header, int feat)
  80 {
  81         clear_bit(feat, header->adds_features);
  82 }
  83
  84 bool perf_header__has_feat(const struct perf_header *header, int feat)
  85 {
  86         return test_bit(feat, header->adds_features);
  87 }
  88
  89 static int __do_write_fd(struct feat_fd *ff, const void *buf, size_t size)
  90 {
  91         ssize_t ret = writen(ff->fd, buf, size);
  92
  93         if (ret != (ssize_t)size)
  94                 return ret < 0 ? (int)ret : -1;
  95         return 0;
  96 }
  97
  98 static int __do_write_buf(struct feat_fd *ff,  const void *buf, size_t size)
  99 {
 100         /* struct perf_event_header::size is u16 */
 101         const size_t max_size = 0xffff - sizeof(struct perf_event_header);
 102         size_t new_size = ff->size;
 103         void *addr;
 104
 105         if (size + ff->offset > max_size)
 106                 return -E2BIG;
 107
 108         while (size > (new_size - ff->offset))
 109                 new_size <<= 1;
 110         new_size = min(max_size, new_size);
 111
 112         if (ff->size < new_size) {
 113                 addr = realloc(ff->buf, new_size);
 114                 if (!addr)
 115                         return -ENOMEM;
 116                 ff->buf = addr;
 117                 ff->size = new_size;
 118         }
 119
 120         memcpy(ff->buf + ff->offset, buf, size);
 121         ff->offset += size;
 122
 123         return 0;
 124 }
 125
 126 /* Return: 0 if succeded, -ERR if failed. */
 127 int do_write(struct feat_fd *ff, const void *buf, size_t size)
 128 {
 129         if (!ff->buf)
 130                 return __do_write_fd(ff, buf, size);
 131         return __do_write_buf(ff, buf, size);
 132 }
 133
 134 /* Return: 0 if succeded, -ERR if failed. */
 135 int write_padded(struct feat_fd *ff, const void *bf,
 136                  size_t count, size_t count_aligned)
 137 {
 138         static const char zero_buf[NAME_ALIGN];
 139         int err = do_write(ff, bf, count);
 140
 141         if (!err)
 142                 err = do_write(ff, zero_buf, count_aligned - count);
 143
 144         return err;
 145 }
 146
 147 #define string_size(str)                                                \
 148         (PERF_ALIGN((strlen(str) + 1), NAME_ALIGN) + sizeof(u32))
 149
 150 /* Return: 0 if succeded, -ERR if failed. */
 151 static int do_write_string(struct feat_fd *ff, const char *str)
 152 {
 153         u32 len, olen;
 154         int ret;
 155
 156         olen = strlen(str) + 1;
 157         len = PERF_ALIGN(olen, NAME_ALIGN);
 158
 159         /* write len, incl. \0 */
 160         ret = do_write(ff, &len, sizeof(len));
 161         if (ret < 0)
 162                 return ret;
 163
 164         return write_padded(ff, str, olen, len);
 165 }
 166
 167 static int __do_read_fd(struct feat_fd *ff, void *addr, ssize_t size)
 168 {
 169         ssize_t ret = readn(ff->fd, addr, size);
 170
 171         if (ret != size)
 172                 return ret < 0 ? (int)ret : -1;
 173         return 0;
 174 }
 175
 176 static int __do_read_buf(struct feat_fd *ff, void *addr, ssize_t size)
 177 {
 178         if (size > (ssize_t)ff->size - ff->offset)
 179                 return -1;
 180
 181         memcpy(addr, ff->buf + ff->offset, size);
 182         ff->offset += size;
 183
 184         return 0;
 185
 186 }
 187
 188 static int __do_read(struct feat_fd *ff, void *addr, ssize_t size)
 189 {
 190         if (!ff->buf)
 191                 return __do_read_fd(ff, addr, size);
 192         return __do_read_buf(ff, addr, size);
 193 }
 194
 195 static int do_read_u32(struct feat_fd *ff, u32 *addr)
 196 {
 197         int ret;
 198
 199         ret = __do_read(ff, addr, sizeof(*addr));
 200         if (ret)
 201                 return ret;
 202
 203         if (ff->ph->needs_swap)
 204                 *addr = bswap_32(*addr);
 205         return 0;
 206 }
 207
 208 static int do_read_u64(struct feat_fd *ff, u64 *addr)
 209 {
 210         int ret;
 211
 212         ret = __do_read(ff, addr, sizeof(*addr));
 213         if (ret)
 214                 return ret;
 215
 216         if (ff->ph->needs_swap)
 217                 *addr = bswap_64(*addr);
 218         return 0;
 219 }
 220
 221 static char *do_read_string(struct feat_fd *ff)
 222 {
 223         u32 len;
 224         char *buf;
 225
 226         if (do_read_u32(ff, &len))
 227                 return NULL;
 228
 229         buf = malloc(len);
 230         if (!buf)
 231                 return NULL;
 232
 233         if (!__do_read(ff, buf, len)) {
 234                 /*
 235                  * strings are padded by zeroes
 236                  * thus the actual strlen of buf
 237                  * may be less than len
 238                  */
 239                 return buf;
 240         }
 241
 242         free(buf);
 243         return NULL;
 244 }
 245
 246 static int write_tracing_data(struct feat_fd *ff,
 247                               struct perf_evlist *evlist)
 248 {
 249         if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
 250                 return -1;
 251
 252         return read_tracing_data(ff->fd, &evlist->entries);
 253 }
 254
 255 static int write_build_id(struct feat_fd *ff,
 256                           struct perf_evlist *evlist __maybe_unused)
 257 {
 258         struct perf_session *session;
 259         int err;
 260
 261         session = container_of(ff->ph, struct perf_session, header);
 262
 263         if (!perf_session__read_build_ids(session, true))
 264                 return -1;
 265
 266         if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
 267                 return -1;
 268
 269         err = perf_session__write_buildid_table(session, ff);
 270         if (err < 0) {
 271                 pr_debug("failed to write buildid table\n");
 272                 return err;
 273         }
 274         perf_session__cache_build_ids(session);
 275
 276         return 0;
 277 }
 278
 279 static int write_hostname(struct feat_fd *ff,
 280                           struct perf_evlist *evlist __maybe_unused)
 281 {
 282         struct utsname uts;
 283         int ret;
 284
 285         ret = uname(&uts);
 286         if (ret < 0)
 287                 return -1;
 288
 289         return do_write_string(ff, uts.nodename);
 290 }
 291
 292 static int write_osrelease(struct feat_fd *ff,
 293                            struct perf_evlist *evlist __maybe_unused)
 294 {
 295         struct utsname uts;
 296         int ret;
 297
 298         ret = uname(&uts);
 299         if (ret < 0)
 300                 return -1;
 301
 302         return do_write_string(ff, uts.release);
 303 }
 304
 305 static int write_arch(struct feat_fd *ff,
 306                       struct perf_evlist *evlist __maybe_unused)
 307 {
 308         struct utsname uts;
 309         int ret;
 310
 311         ret = uname(&uts);
 312         if (ret < 0)
 313                 return -1;
 314
 315         return do_write_string(ff, uts.machine);
 316 }
 317
 318 static int write_version(struct feat_fd *ff,
 319                          struct perf_evlist *evlist __maybe_unused)
 320 {
 321         return do_write_string(ff, perf_version_string);
 322 }
 323
 324 static int __write_cpudesc(struct feat_fd *ff, const char *cpuinfo_proc)
 325 {
 326         FILE *file;
 327         char *buf = NULL;
 328         char *s, *p;
 329         const char *search = cpuinfo_proc;
 330         size_t len = 0;
 331         int ret = -1;
 332
 333         if (!search)
 334                 return -1;
 335
 336         file = fopen("/proc/cpuinfo", "r");
 337         if (!file)
 338                 return -1;
 339
 340         while (getline(&buf, &len, file) > 0) {
 341                 ret = strncmp(buf, search, strlen(search));
 342                 if (!ret)
 343                         break;
 344         }
 345
 346         if (ret) {
 347                 ret = -1;
 348                 goto done;
 349         }
 350
 351         s = buf;
 352
 353         p = strchr(buf, ':');
 354         if (p && *(p+1) == ' ' && *(p+2))
 355                 s = p + 2;
 356         p = strchr(s, '\n');
 357         if (p)
 358                 *p = '\0';
 359
 360         /* squash extra space characters (branding string) */
 361         p = s;
 362         while (*p) {
 363                 if (isspace(*p)) {
 364                         char *r = p + 1;
 365                         char *q = r;
 366                         *p = ' ';
 367                         while (*q && isspace(*q))
 368                                 q++;
 369                         if (q != (p+1))
 370                                 while ((*r++ = *q++));
 371                 }
 372                 p++;
 373         }
 374         ret = do_write_string(ff, s);
 375 done:
 376         free(buf);
 377         fclose(file);
 378         return ret;
 379 }
 380
 381 static int write_cpudesc(struct feat_fd *ff,
 382                        struct perf_evlist *evlist __maybe_unused)
 383 {
 384         const char *cpuinfo_procs[] = CPUINFO_PROC;
 385         unsigned int i;
 386
 387         for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
 388                 int ret;
 389                 ret = __write_cpudesc(ff, cpuinfo_procs[i]);
 390                 if (ret >= 0)
 391                         return ret;
 392         }
 393         return -1;
 394 }
 395
 396
 397 static int write_nrcpus(struct feat_fd *ff,
 398                         struct perf_evlist *evlist __maybe_unused)
 399 {
 400         long nr;
 401         u32 nrc, nra;
 402         int ret;
 403
 404         nrc = cpu__max_present_cpu();
 405
 406         nr = sysconf(_SC_NPROCESSORS_ONLN);
 407         if (nr < 0)
 408                 return -1;
 409
 410         nra = (u32)(nr & UINT_MAX);
 411
 412         ret = do_write(ff, &nrc, sizeof(nrc));
 413         if (ret < 0)
 414                 return ret;
 415
 416         return do_write(ff, &nra, sizeof(nra));
 417 }
 418
 419 static int write_event_desc(struct feat_fd *ff,
 420                             struct perf_evlist *evlist)
 421 {
 422         struct perf_evsel *evsel;
 423         u32 nre, nri, sz;
 424         int ret;
 425
 426         nre = evlist->nr_entries;
 427
 428         /*
 429          * write number of events
 430          */
 431         ret = do_write(ff, &nre, sizeof(nre));
 432         if (ret < 0)
 433                 return ret;
 434
 435         /*
 436          * size of perf_event_attr struct
 437          */
 438         sz = (u32)sizeof(evsel->attr);
 439         ret = do_write(ff, &sz, sizeof(sz));
 440         if (ret < 0)
 441                 return ret;
 442
 443         evlist__for_each_entry(evlist, evsel) {
 444                 ret = do_write(ff, &evsel->attr, sz);
 445                 if (ret < 0)
 446                         return ret;
 447                 /*
 448                  * write number of unique id per event
 449                  * there is one id per instance of an event
 450                  *
 451                  * copy into an nri to be independent of the
 452                  * type of ids,
 453                  */
 454                 nri = evsel->ids;
 455                 ret = do_write(ff, &nri, sizeof(nri));
 456                 if (ret < 0)
 457                         return ret;
 458
 459                 /*
 460                  * write event string as passed on cmdline
 461                  */
 462                 ret = do_write_string(ff, perf_evsel__name(evsel));
 463                 if (ret < 0)
 464                         return ret;
 465                 /*
 466                  * write unique ids for this event
 467                  */
 468                 ret = do_write(ff, evsel->id, evsel->ids * sizeof(u64));
 469                 if (ret < 0)
 470                         return ret;
 471         }
 472         return 0;
 473 }
 474
 475 static int write_cmdline(struct feat_fd *ff,
 476                          struct perf_evlist *evlist __maybe_unused)
 477 {
 478         char buf[MAXPATHLEN];
 479         u32 n;
 480         int i, ret;
 481
 482         /* actual path to perf binary */
 483         ret = readlink("/proc/self/exe", buf, sizeof(buf) - 1);
 484         if (ret <= 0)
 485                 return -1;
 486
 487         /* readlink() does not add null termination */
 488         buf[ret] = '\0';
 489
 490         /* account for binary path */
 491         n = perf_env.nr_cmdline + 1;
 492
 493         ret = do_write(ff, &n, sizeof(n));
 494         if (ret < 0)
 495                 return ret;
 496
 497         ret = do_write_string(ff, buf);
 498         if (ret < 0)
 499                 return ret;
 500
 501         for (i = 0 ; i < perf_env.nr_cmdline; i++) {
 502                 ret = do_write_string(ff, perf_env.cmdline_argv[i]);
 503                 if (ret < 0)
 504                         return ret;
 505         }
 506         return 0;
 507 }
 508
 509 #define CORE_SIB_FMT \
 510         "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
 511 #define THRD_SIB_FMT \
 512         "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
 513
 514 struct cpu_topo {
 515         u32 cpu_nr;
 516         u32 core_sib;
 517         u32 thread_sib;
 518         char **core_siblings;
 519         char **thread_siblings;
 520 };
 521
 522 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
 523 {
 524         FILE *fp;
 525         char filename[MAXPATHLEN];
 526         char *buf = NULL, *p;
 527         size_t len = 0;
 528         ssize_t sret;
 529         u32 i = 0;
 530         int ret = -1;
 531
 532         sprintf(filename, CORE_SIB_FMT, cpu);
 533         fp = fopen(filename, "r");
 534         if (!fp)
 535                 goto try_threads;
 536
 537         sret = getline(&buf, &len, fp);
 538         fclose(fp);
 539         if (sret <= 0)
 540                 goto try_threads;
 541
 542         p = strchr(buf, '\n');
 543         if (p)
 544                 *p = '\0';
 545
 546         for (i = 0; i < tp->core_sib; i++) {
 547                 if (!strcmp(buf, tp->core_siblings[i]))
 548                         break;
 549         }
 550         if (i == tp->core_sib) {
 551                 tp->core_siblings[i] = buf;
 552                 tp->core_sib++;
 553                 buf = NULL;
 554                 len = 0;
 555         }
 556         ret = 0;
 557
 558 try_threads:
 559         sprintf(filename, THRD_SIB_FMT, cpu);
 560         fp = fopen(filename, "r");
 561         if (!fp)
 562                 goto done;
 563
 564         if (getline(&buf, &len, fp) <= 0)
 565                 goto done;
 566
 567         p = strchr(buf, '\n');
 568         if (p)
 569                 *p = '\0';
 570
 571         for (i = 0; i < tp->thread_sib; i++) {
 572                 if (!strcmp(buf, tp->thread_siblings[i]))
 573                         break;
 574         }
 575         if (i == tp->thread_sib) {
 576                 tp->thread_siblings[i] = buf;
 577                 tp->thread_sib++;
 578                 buf = NULL;
 579         }
 580         ret = 0;
 581 done:
 582         if(fp)
 583                 fclose(fp);
 584         free(buf);
 585         return ret;
 586 }
 587
 588 static void free_cpu_topo(struct cpu_topo *tp)
 589 {
 590         u32 i;
 591
 592         if (!tp)
 593                 return;
 594
 595         for (i = 0 ; i < tp->core_sib; i++)
 596                 zfree(&tp->core_siblings[i]);
 597
 598         for (i = 0 ; i < tp->thread_sib; i++)
 599                 zfree(&tp->thread_siblings[i]);
 600
 601         free(tp);
 602 }
 603
 604 static struct cpu_topo *build_cpu_topology(void)
 605 {
 606         struct cpu_topo *tp = NULL;
 607         void *addr;
 608         u32 nr, i;
 609         size_t sz;
 610         long ncpus;
 611         int ret = -1;
 612         struct cpu_map *map;
 613
 614         ncpus = cpu__max_present_cpu();
 615
 616         /* build online CPU map */
 617         map = cpu_map__new(NULL);
 618         if (map == NULL) {
 619                 pr_debug("failed to get system cpumap\n");
 620                 return NULL;
 621         }
 622
 623         nr = (u32)(ncpus & UINT_MAX);
 624
 625         sz = nr * sizeof(char *);
 626         addr = calloc(1, sizeof(*tp) + 2 * sz);
 627         if (!addr)
 628                 goto out_free;
 629
 630         tp = addr;
 631         tp->cpu_nr = nr;
 632         addr += sizeof(*tp);
 633         tp->core_siblings = addr;
 634         addr += sz;
 635         tp->thread_siblings = addr;
 636
 637         for (i = 0; i < nr; i++) {
 638                 if (!cpu_map__has(map, i))
 639                         continue;
 640
 641                 ret = build_cpu_topo(tp, i);
 642                 if (ret < 0)
 643                         break;
 644         }
 645
 646 out_free:
 647         cpu_map__put(map);
 648         if (ret) {
 649                 free_cpu_topo(tp);
 650                 tp = NULL;
 651         }
 652         return tp;
 653 }
 654
 655 static int write_cpu_topology(struct feat_fd *ff,
 656                               struct perf_evlist *evlist __maybe_unused)
 657 {
 658         struct cpu_topo *tp;
 659         u32 i;
 660         int ret, j;
 661
 662         tp = build_cpu_topology();
 663         if (!tp)
 664                 return -1;
 665
 666         ret = do_write(ff, &tp->core_sib, sizeof(tp->core_sib));
 667         if (ret < 0)
 668                 goto done;
 669
 670         for (i = 0; i < tp->core_sib; i++) {
 671                 ret = do_write_string(ff, tp->core_siblings[i]);
 672                 if (ret < 0)
 673                         goto done;
 674         }
 675         ret = do_write(ff, &tp->thread_sib, sizeof(tp->thread_sib));
 676         if (ret < 0)
 677                 goto done;
 678
 679         for (i = 0; i < tp->thread_sib; i++) {
 680                 ret = do_write_string(ff, tp->thread_siblings[i]);
 681                 if (ret < 0)
 682                         break;
 683         }
 684
 685         ret = perf_env__read_cpu_topology_map(&perf_env);
 686         if (ret < 0)
 687                 goto done;
 688
 689         for (j = 0; j < perf_env.nr_cpus_avail; j++) {
 690                 ret = do_write(ff, &perf_env.cpu[j].core_id,
 691                                sizeof(perf_env.cpu[j].core_id));
 692                 if (ret < 0)
 693                         return ret;
 694                 ret = do_write(ff, &perf_env.cpu[j].socket_id,
 695                                sizeof(perf_env.cpu[j].socket_id));
 696                 if (ret < 0)
 697                         return ret;
 698         }
 699 done:
 700         free_cpu_topo(tp);
 701         return ret;
 702 }
 703
 704
 705
 706 static int write_total_mem(struct feat_fd *ff,
 707                            struct perf_evlist *evlist __maybe_unused)
 708 {
 709         char *buf = NULL;
 710         FILE *fp;
 711         size_t len = 0;
 712         int ret = -1, n;
 713         uint64_t mem;
 714
 715         fp = fopen("/proc/meminfo", "r");
 716         if (!fp)
 717                 return -1;
 718
 719         while (getline(&buf, &len, fp) > 0) {
 720                 ret = strncmp(buf, "MemTotal:", 9);
 721                 if (!ret)
 722                         break;
 723         }
 724         if (!ret) {
 725                 n = sscanf(buf, "%*s %"PRIu64, &mem);
 726                 if (n == 1)
 727                         ret = do_write(ff, &mem, sizeof(mem));
 728         } else
 729                 ret = -1;
 730         free(buf);
 731         fclose(fp);
 732         return ret;
 733 }
 734
 735 static int write_topo_node(struct feat_fd *ff, int node)
 736 {
 737         char str[MAXPATHLEN];
 738         char field[32];
 739         char *buf = NULL, *p;
 740         size_t len = 0;
 741         FILE *fp;
 742         u64 mem_total, mem_free, mem;
 743         int ret = -1;
 744
 745         sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
 746         fp = fopen(str, "r");
 747         if (!fp)
 748                 return -1;
 749
 750         while (getline(&buf, &len, fp) > 0) {
 751                 /* skip over invalid lines */
 752                 if (!strchr(buf, ':'))
 753                         continue;
 754                 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
 755                         goto done;
 756                 if (!strcmp(field, "MemTotal:"))
 757                         mem_total = mem;
 758                 if (!strcmp(field, "MemFree:"))
 759                         mem_free = mem;
 760         }
 761
 762         fclose(fp);
 763         fp = NULL;
 764
 765         ret = do_write(ff, &mem_total, sizeof(u64));
 766         if (ret)
 767                 goto done;
 768
 769         ret = do_write(ff, &mem_free, sizeof(u64));
 770         if (ret)
 771                 goto done;
 772
 773         ret = -1;
 774         sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
 775
 776         fp = fopen(str, "r");
 777         if (!fp)
 778                 goto done;
 779
 780         if (getline(&buf, &len, fp) <= 0)
 781                 goto done;
 782
 783         p = strchr(buf, '\n');
 784         if (p)
 785                 *p = '\0';
 786
 787         ret = do_write_string(ff, buf);
 788 done:
 789         free(buf);
 790         if (fp)
 791                 fclose(fp);
 792         return ret;
 793 }
 794
 795 static int write_numa_topology(struct feat_fd *ff,
 796                                struct perf_evlist *evlist __maybe_unused)
 797 {
 798         char *buf = NULL;
 799         size_t len = 0;
 800         FILE *fp;
 801         struct cpu_map *node_map = NULL;
 802         char *c;
 803         u32 nr, i, j;
 804         int ret = -1;
 805
 806         fp = fopen("/sys/devices/system/node/online", "r");
 807         if (!fp)
 808                 return -1;
 809
 810         if (getline(&buf, &len, fp) <= 0)
 811                 goto done;
 812
 813         c = strchr(buf, '\n');
 814         if (c)
 815                 *c = '\0';
 816
 817         node_map = cpu_map__new(buf);
 818         if (!node_map)
 819                 goto done;
 820
 821         nr = (u32)node_map->nr;
 822
 823         ret = do_write(ff, &nr, sizeof(nr));
 824         if (ret < 0)
 825                 goto done;
 826
 827         for (i = 0; i < nr; i++) {
 828                 j = (u32)node_map->map[i];
 829                 ret = do_write(ff, &j, sizeof(j));
 830                 if (ret < 0)
 831                         break;
 832
 833                 ret = write_topo_node(ff, i);
 834                 if (ret < 0)
 835                         break;
 836         }
 837 done:
 838         free(buf);
 839         fclose(fp);
 840         cpu_map__put(node_map);
 841         return ret;
 842 }
 843
 844 /*
 845  * File format:
 846  *
 847  * struct pmu_mappings {
 848  *      u32     pmu_num;
 849  *      struct pmu_map {
 850  *              u32     type;
 851  *              char    name[];
 852  *      }[pmu_num];
 853  * };
 854  */
 855
 856 static int write_pmu_mappings(struct feat_fd *ff,
 857                               struct perf_evlist *evlist __maybe_unused)
 858 {
 859         struct perf_pmu *pmu = NULL;
 860         u32 pmu_num = 0;
 861         int ret;
 862
 863         /*
 864          * Do a first pass to count number of pmu to avoid lseek so this
 865          * works in pipe mode as well.
 866          */
 867         while ((pmu = perf_pmu__scan(pmu))) {
 868                 if (!pmu->name)
 869                         continue;
 870                 pmu_num++;
 871         }
 872
 873         ret = do_write(ff, &pmu_num, sizeof(pmu_num));
 874         if (ret < 0)
 875                 return ret;
 876
 877         while ((pmu = perf_pmu__scan(pmu))) {
 878                 if (!pmu->name)
 879                         continue;
 880
 881                 ret = do_write(ff, &pmu->type, sizeof(pmu->type));
 882                 if (ret < 0)
 883                         return ret;
 884
 885                 ret = do_write_string(ff, pmu->name);
 886                 if (ret < 0)
 887                         return ret;
 888         }
 889
 890         return 0;
 891 }
 892
 893 /*
 894  * File format:
 895  *
 896  * struct group_descs {
 897  *      u32     nr_groups;
 898  *      struct group_desc {
 899  *              char    name[];
 900  *              u32     leader_idx;
 901  *              u32     nr_members;
 902  *      }[nr_groups];
 903  * };
 904  */
 905 static int write_group_desc(struct feat_fd *ff,
 906                             struct perf_evlist *evlist)
 907 {
 908         u32 nr_groups = evlist->nr_groups;
 909         struct perf_evsel *evsel;
 910         int ret;
 911
 912         ret = do_write(ff, &nr_groups, sizeof(nr_groups));
 913         if (ret < 0)
 914                 return ret;
 915
 916         evlist__for_each_entry(evlist, evsel) {
 917                 if (perf_evsel__is_group_leader(evsel) &&
 918                     evsel->nr_members > 1) {
 919                         const char *name = evsel->group_name ?: "{anon_group}";
 920                         u32 leader_idx = evsel->idx;
 921                         u32 nr_members = evsel->nr_members;
 922
 923                         ret = do_write_string(ff, name);
 924                         if (ret < 0)
 925                                 return ret;
 926
 927                         ret = do_write(ff, &leader_idx, sizeof(leader_idx));
 928                         if (ret < 0)
 929                                 return ret;
 930
 931                         ret = do_write(ff, &nr_members, sizeof(nr_members));
 932                         if (ret < 0)
 933                                 return ret;
 934                 }
 935         }
 936         return 0;
 937 }
 938
 939 /*
 940  * default get_cpuid(): nothing gets recorded
 941  * actual implementation must be in arch/$(SRCARCH)/util/header.c
 942  */
 943 int __weak get_cpuid(char *buffer __maybe_unused, size_t sz __maybe_unused)
 944 {
 945         return -1;
 946 }
 947
 948 static int write_cpuid(struct feat_fd *ff,
 949                        struct perf_evlist *evlist __maybe_unused)
 950 {
 951         char buffer[64];
 952         int ret;
 953
 954         ret = get_cpuid(buffer, sizeof(buffer));
 955         if (!ret)
 956                 goto write_it;
 957
 958         return -1;
 959 write_it:
 960         return do_write_string(ff, buffer);
 961 }
 962
 963 static int write_branch_stack(struct feat_fd *ff __maybe_unused,
 964                               struct perf_evlist *evlist __maybe_unused)
 965 {
 966         return 0;
 967 }
 968
 969 static int write_auxtrace(struct feat_fd *ff,
 970                           struct perf_evlist *evlist __maybe_unused)
 971 {
 972         struct perf_session *session;
 973         int err;
 974
 975         if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
 976                 return -1;
 977
 978         session = container_of(ff->ph, struct perf_session, header);
 979
 980         err = auxtrace_index__write(ff->fd, &session->auxtrace_index);
 981         if (err < 0)
 982                 pr_err("Failed to write auxtrace index\n");
 983         return err;
 984 }
 985
 986 static int cpu_cache_level__sort(const void *a, const void *b)
 987 {
 988         struct cpu_cache_level *cache_a = (struct cpu_cache_level *)a;
 989         struct cpu_cache_level *cache_b = (struct cpu_cache_level *)b;
 990
 991         return cache_a->level - cache_b->level;
 992 }
 993
 994 static bool cpu_cache_level__cmp(struct cpu_cache_level *a, struct cpu_cache_level *b)
 995 {
 996         if (a->level != b->level)
 997                 return false;
 998
 999         if (a->line_size != b->line_size)
1000                 return false;
1001
1002         if (a->sets != b->sets)
1003                 return false;
1004
1005         if (a->ways != b->ways)
1006                 return false;
1007
1008         if (strcmp(a->type, b->type))
1009                 return false;
1010
1011         if (strcmp(a->size, b->size))
1012                 return false;
1013
1014         if (strcmp(a->map, b->map))
1015                 return false;
1016
1017         return true;
1018 }
1019
1020 static int cpu_cache_level__read(struct cpu_cache_level *cache, u32 cpu, u16 level)
1021 {
1022         char path[PATH_MAX], file[PATH_MAX];
1023         struct stat st;
1024         size_t len;
1025
1026         scnprintf(path, PATH_MAX, "devices/system/cpu/cpu%d/cache/index%d/", cpu, level);
1027         scnprintf(file, PATH_MAX, "%s/%s", sysfs__mountpoint(), path);
1028
1029         if (stat(file, &st))
1030                 return 1;
1031
1032         scnprintf(file, PATH_MAX, "%s/level", path);
1033         if (sysfs__read_int(file, (int *) &cache->level))
1034                 return -1;
1035
1036         scnprintf(file, PATH_MAX, "%s/coherency_line_size", path);
1037         if (sysfs__read_int(file, (int *) &cache->line_size))
1038                 return -1;
1039
1040         scnprintf(file, PATH_MAX, "%s/number_of_sets", path);
1041         if (sysfs__read_int(file, (int *) &cache->sets))
1042                 return -1;
1043
1044         scnprintf(file, PATH_MAX, "%s/ways_of_associativity", path);
1045         if (sysfs__read_int(file, (int *) &cache->ways))
1046                 return -1;
1047
1048         scnprintf(file, PATH_MAX, "%s/type", path);
1049         if (sysfs__read_str(file, &cache->type, &len))
1050                 return -1;
1051
1052         cache->type[len] = 0;
1053         cache->type = rtrim(cache->type);
1054
1055         scnprintf(file, PATH_MAX, "%s/size", path);
1056         if (sysfs__read_str(file, &cache->size, &len)) {
1057                 free(cache->type);
1058                 return -1;
1059         }
1060
1061         cache->size[len] = 0;
1062         cache->size = rtrim(cache->size);
1063
1064         scnprintf(file, PATH_MAX, "%s/shared_cpu_list", path);
1065         if (sysfs__read_str(file, &cache->map, &len)) {
1066                 free(cache->map);
1067                 free(cache->type);
1068                 return -1;
1069         }
1070
1071         cache->map[len] = 0;
1072         cache->map = rtrim(cache->map);
1073         return 0;
1074 }
1075
1076 static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c)
1077 {
1078         fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map);
1079 }
1080
1081 static int build_caches(struct cpu_cache_level caches[], u32 size, u32 *cntp)
1082 {
1083         u32 i, cnt = 0;
1084         long ncpus;
1085         u32 nr, cpu;
1086         u16 level;
1087
1088         ncpus = sysconf(_SC_NPROCESSORS_CONF);
1089         if (ncpus < 0)
1090                 return -1;
1091
1092         nr = (u32)(ncpus & UINT_MAX);
1093
1094         for (cpu = 0; cpu < nr; cpu++) {
1095                 for (level = 0; level < 10; level++) {
1096                         struct cpu_cache_level c;
1097                         int err;
1098
1099                         err = cpu_cache_level__read(&c, cpu, level);
1100                         if (err < 0)
1101                                 return err;
1102
1103                         if (err == 1)
1104                                 break;
1105
1106                         for (i = 0; i < cnt; i++) {
1107                                 if (cpu_cache_level__cmp(&c, &caches[i]))
1108                                         break;
1109                         }
1110
1111                         if (i == cnt)
1112                                 caches[cnt++] = c;
1113                         else
1114                                 cpu_cache_level__free(&c);
1115
1116                         if (WARN_ONCE(cnt == size, "way too many cpu caches.."))
1117                                 goto out;
1118                 }
1119         }
1120  out:
1121         *cntp = cnt;
1122         return 0;
1123 }
1124
1125 #define MAX_CACHES 2000
1126
1127 static int write_cache(struct feat_fd *ff,
1128                        struct perf_evlist *evlist __maybe_unused)
1129 {
1130         struct cpu_cache_level caches[MAX_CACHES];
1131         u32 cnt = 0, i, version = 1;
1132         int ret;
1133
1134         ret = build_caches(caches, MAX_CACHES, &cnt);
1135         if (ret)
1136                 goto out;
1137
1138         qsort(&caches, cnt, sizeof(struct cpu_cache_level), cpu_cache_level__sort);
1139
1140         ret = do_write(ff, &version, sizeof(u32));
1141         if (ret < 0)
1142                 goto out;
1143
1144         ret = do_write(ff, &cnt, sizeof(u32));
1145         if (ret < 0)
1146                 goto out;
1147
1148         for (i = 0; i < cnt; i++) {
1149                 struct cpu_cache_level *c = &caches[i];
1150
1151                 #define _W(v)                                   \
1152                         ret = do_write(ff, &c->v, sizeof(u32)); \
1153                         if (ret < 0)                            \
1154                                 goto out;
1155
1156                 _W(level)
1157                 _W(line_size)
1158                 _W(sets)
1159                 _W(ways)
1160                 #undef _W
1161
1162                 #define _W(v)                                           \
1163                         ret = do_write_string(ff, (const char *) c->v); \
1164                         if (ret < 0)                                    \
1165                                 goto out;
1166
1167                 _W(type)
1168                 _W(size)
1169                 _W(map)
1170                 #undef _W
1171         }
1172
1173 out:
1174         for (i = 0; i < cnt; i++)
1175                 cpu_cache_level__free(&caches[i]);
1176         return ret;
1177 }
1178
1179 static int write_stat(struct feat_fd *ff __maybe_unused,
1180                       struct perf_evlist *evlist __maybe_unused)
1181 {
1182         return 0;
1183 }
1184
1185 static void print_hostname(struct feat_fd *ff, FILE *fp)
1186 {
1187         fprintf(fp, "# hostname : %s\n", ff->ph->env.hostname);
1188 }
1189
1190 static void print_osrelease(struct feat_fd *ff, FILE *fp)
1191 {
1192         fprintf(fp, "# os release : %s\n", ff->ph->env.os_release);
1193 }
1194
1195 static void print_arch(struct feat_fd *ff, FILE *fp)
1196 {
1197         fprintf(fp, "# arch : %s\n", ff->ph->env.arch);
1198 }
1199
1200 static void print_cpudesc(struct feat_fd *ff, FILE *fp)
1201 {
1202         fprintf(fp, "# cpudesc : %s\n", ff->ph->env.cpu_desc);
1203 }
1204
1205 static void print_nrcpus(struct feat_fd *ff, FILE *fp)
1206 {
1207         fprintf(fp, "# nrcpus online : %u\n", ff->ph->env.nr_cpus_online);
1208         fprintf(fp, "# nrcpus avail : %u\n", ff->ph->env.nr_cpus_avail);
1209 }
1210
1211 static void print_version(struct feat_fd *ff, FILE *fp)
1212 {
1213         fprintf(fp, "# perf version : %s\n", ff->ph->env.version);
1214 }
1215
1216 static void print_cmdline(struct feat_fd *ff, FILE *fp)
1217 {
1218         int nr, i;
1219
1220         nr = ff->ph->env.nr_cmdline;
1221
1222         fprintf(fp, "# cmdline : ");
1223
1224         for (i = 0; i < nr; i++)
1225                 fprintf(fp, "%s ", ff->ph->env.cmdline_argv[i]);
1226         fputc('\n', fp);
1227 }
1228
1229 static void print_cpu_topology(struct feat_fd *ff, FILE *fp)
1230 {
1231         struct perf_header *ph = ff->ph;
1232         int cpu_nr = ph->env.nr_cpus_avail;
1233         int nr, i;
1234         char *str;
1235
1236         nr = ph->env.nr_sibling_cores;
1237         str = ph->env.sibling_cores;
1238
1239         for (i = 0; i < nr; i++) {
1240                 fprintf(fp, "# sibling cores   : %s\n", str);
1241                 str += strlen(str) + 1;
1242         }
1243
1244         nr = ph->env.nr_sibling_threads;
1245         str = ph->env.sibling_threads;
1246
1247         for (i = 0; i < nr; i++) {
1248                 fprintf(fp, "# sibling threads : %s\n", str);
1249                 str += strlen(str) + 1;
1250         }
1251
1252         if (ph->env.cpu != NULL) {
1253                 for (i = 0; i < cpu_nr; i++)
1254                         fprintf(fp, "# CPU %d: Core ID %d, Socket ID %d\n", i,
1255                                 ph->env.cpu[i].core_id, ph->env.cpu[i].socket_id);
1256         } else
1257                 fprintf(fp, "# Core ID and Socket ID information is not available\n");
1258 }
1259
1260 static void free_event_desc(struct perf_evsel *events)
1261 {
1262         struct perf_evsel *evsel;
1263
1264         if (!events)
1265                 return;
1266
1267         for (evsel = events; evsel->attr.size; evsel++) {
1268                 zfree(&evsel->name);
1269                 zfree(&evsel->id);
1270         }
1271
1272         free(events);
1273 }
1274
1275 static struct perf_evsel *read_event_desc(struct feat_fd *ff)
1276 {
1277         struct perf_evsel *evsel, *events = NULL;
1278         u64 *id;
1279         void *buf = NULL;
1280         u32 nre, sz, nr, i, j;
1281         size_t msz;
1282
1283         /* number of events */
1284         if (do_read_u32(ff, &nre))
1285                 goto error;
1286
1287         if (do_read_u32(ff, &sz))
1288                 goto error;
1289
1290         /* buffer to hold on file attr struct */
1291         buf = malloc(sz);
1292         if (!buf)
1293                 goto error;
1294
1295         /* the last event terminates with evsel->attr.size == 0: */
1296         events = calloc(nre + 1, sizeof(*events));
1297         if (!events)
1298                 goto error;
1299
1300         msz = sizeof(evsel->attr);
1301         if (sz < msz)
1302                 msz = sz;
1303
1304         for (i = 0, evsel = events; i < nre; evsel++, i++) {
1305                 evsel->idx = i;
1306
1307                 /*
1308                  * must read entire on-file attr struct to
1309                  * sync up with layout.
1310                  */
1311                 if (__do_read(ff, buf, sz))
1312                         goto error;
1313
1314                 if (ff->ph->needs_swap)
1315                         perf_event__attr_swap(buf);
1316
1317                 memcpy(&evsel->attr, buf, msz);
1318
1319                 if (do_read_u32(ff, &nr))
1320                         goto error;
1321
1322                 if (ff->ph->needs_swap)
1323                         evsel->needs_swap = true;
1324
1325                 evsel->name = do_read_string(ff);
1326                 if (!evsel->name)
1327                         goto error;
1328
1329                 if (!nr)
1330                         continue;
1331
1332                 id = calloc(nr, sizeof(*id));
1333                 if (!id)
1334                         goto error;
1335                 evsel->ids = nr;
1336                 evsel->id = id;
1337
1338                 for (j = 0 ; j < nr; j++) {
1339                         if (do_read_u64(ff, id))
1340                                 goto error;
1341                         id++;
1342                 }
1343         }
1344 out:
1345         free(buf);
1346         return events;
1347 error:
1348         free_event_desc(events);
1349         events = NULL;
1350         goto out;
1351 }
1352
1353 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1354                                 void *priv __maybe_unused)
1355 {
1356         return fprintf(fp, ", %s = %s", name, val);
1357 }
1358
1359 static void print_event_desc(struct feat_fd *ff, FILE *fp)
1360 {
1361         struct perf_evsel *evsel, *events;
1362         u32 j;
1363         u64 *id;
1364
1365         if (ff->events)
1366                 events = ff->events;
1367         else
1368                 events = read_event_desc(ff);
1369
1370         if (!events) {
1371                 fprintf(fp, "# event desc: not available or unable to read\n");
1372                 return;
1373         }
1374
1375         for (evsel = events; evsel->attr.size; evsel++) {
1376                 fprintf(fp, "# event : name = %s, ", evsel->name);
1377
1378                 if (evsel->ids) {
1379                         fprintf(fp, ", id = {");
1380                         for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1381                                 if (j)
1382                                         fputc(',', fp);
1383                                 fprintf(fp, " %"PRIu64, *id);
1384                         }
1385                         fprintf(fp, " }");
1386                 }
1387
1388                 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1389
1390                 fputc('\n', fp);
1391         }
1392
1393         free_event_desc(events);
1394         ff->events = NULL;
1395 }
1396
1397 static void print_total_mem(struct feat_fd *ff, FILE *fp)
1398 {
1399         fprintf(fp, "# total memory : %llu kB\n", ff->ph->env.total_mem);
1400 }
1401
1402 static void print_numa_topology(struct feat_fd *ff, FILE *fp)
1403 {
1404         int i;
1405         struct numa_node *n;
1406
1407         for (i = 0; i < ff->ph->env.nr_numa_nodes; i++) {
1408                 n = &ff->ph->env.numa_nodes[i];
1409
1410                 fprintf(fp, "# node%u meminfo  : total = %"PRIu64" kB,"
1411                             " free = %"PRIu64" kB\n",
1412                         n->node, n->mem_total, n->mem_free);
1413
1414                 fprintf(fp, "# node%u cpu list : ", n->node);
1415                 cpu_map__fprintf(n->map, fp);
1416         }
1417 }
1418
1419 static void print_cpuid(struct feat_fd *ff, FILE *fp)
1420 {
1421         fprintf(fp, "# cpuid : %s\n", ff->ph->env.cpuid);
1422 }
1423
1424 static void print_branch_stack(struct feat_fd *ff __maybe_unused, FILE *fp)
1425 {
1426         fprintf(fp, "# contains samples with branch stack\n");
1427 }
1428
1429 static void print_auxtrace(struct feat_fd *ff __maybe_unused, FILE *fp)
1430 {
1431         fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1432 }
1433
1434 static void print_stat(struct feat_fd *ff __maybe_unused, FILE *fp)
1435 {
1436         fprintf(fp, "# contains stat data\n");
1437 }
1438
1439 static void print_cache(struct feat_fd *ff, FILE *fp __maybe_unused)
1440 {
1441         int i;
1442
1443         fprintf(fp, "# CPU cache info:\n");
1444         for (i = 0; i < ff->ph->env.caches_cnt; i++) {
1445                 fprintf(fp, "#  ");
1446                 cpu_cache_level__fprintf(fp, &ff->ph->env.caches[i]);
1447         }
1448 }
1449
1450 static void print_pmu_mappings(struct feat_fd *ff, FILE *fp)
1451 {
1452         const char *delimiter = "# pmu mappings: ";
1453         char *str, *tmp;
1454         u32 pmu_num;
1455         u32 type;
1456
1457         pmu_num = ff->ph->env.nr_pmu_mappings;
1458         if (!pmu_num) {
1459                 fprintf(fp, "# pmu mappings: not available\n");
1460                 return;
1461         }
1462
1463         str = ff->ph->env.pmu_mappings;
1464
1465         while (pmu_num) {
1466                 type = strtoul(str, &tmp, 0);
1467                 if (*tmp != ':')
1468                         goto error;
1469
1470                 str = tmp + 1;
1471                 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1472
1473                 delimiter = ", ";
1474                 str += strlen(str) + 1;
1475                 pmu_num--;
1476         }
1477
1478         fprintf(fp, "\n");
1479
1480         if (!pmu_num)
1481                 return;
1482 error:
1483         fprintf(fp, "# pmu mappings: unable to read\n");
1484 }
1485
1486 static void print_group_desc(struct feat_fd *ff, FILE *fp)
1487 {
1488         struct perf_session *session;
1489         struct perf_evsel *evsel;
1490         u32 nr = 0;
1491
1492         session = container_of(ff->ph, struct perf_session, header);
1493
1494         evlist__for_each_entry(session->evlist, evsel) {
1495                 if (perf_evsel__is_group_leader(evsel) &&
1496                     evsel->nr_members > 1) {
1497                         fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1498                                 perf_evsel__name(evsel));
1499
1500                         nr = evsel->nr_members - 1;
1501                 } else if (nr) {
1502                         fprintf(fp, ",%s", perf_evsel__name(evsel));
1503
1504                         if (--nr == 0)
1505                                 fprintf(fp, "}\n");
1506                 }
1507         }
1508 }
1509
1510 static int __event_process_build_id(struct build_id_event *bev,
1511                                     char *filename,
1512                                     struct perf_session *session)
1513 {
1514         int err = -1;
1515         struct machine *machine;
1516         u16 cpumode;
1517         struct dso *dso;
1518         enum dso_kernel_type dso_type;
1519
1520         machine = perf_session__findnew_machine(session, bev->pid);
1521         if (!machine)
1522                 goto out;
1523
1524         cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1525
1526         switch (cpumode) {
1527         case PERF_RECORD_MISC_KERNEL:
1528                 dso_type = DSO_TYPE_KERNEL;
1529                 break;
1530         case PERF_RECORD_MISC_GUEST_KERNEL:
1531                 dso_type = DSO_TYPE_GUEST_KERNEL;
1532                 break;
1533         case PERF_RECORD_MISC_USER:
1534         case PERF_RECORD_MISC_GUEST_USER:
1535                 dso_type = DSO_TYPE_USER;
1536                 break;
1537         default:
1538                 goto out;
1539         }
1540
1541         dso = machine__findnew_dso(machine, filename);
1542         if (dso != NULL) {
1543                 char sbuild_id[SBUILD_ID_SIZE];
1544
1545                 dso__set_build_id(dso, &bev->build_id);
1546
1547                 if (dso_type != DSO_TYPE_USER) {
1548                         struct kmod_path m = { .name = NULL, };
1549
1550                         if (!kmod_path__parse_name(&m, filename) && m.kmod)
1551                                 dso__set_module_info(dso, &m, machine);
1552                         else
1553                                 dso->kernel = dso_type;
1554
1555                         free(m.name);
1556                 }
1557
1558                 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1559                                   sbuild_id);
1560                 pr_debug("build id event received for %s: %s\n",
1561                          dso->long_name, sbuild_id);
1562                 dso__put(dso);
1563         }
1564
1565         err = 0;
1566 out:
1567         return err;
1568 }
1569
1570 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1571                                                  int input, u64 offset, u64 size)
1572 {
1573         struct perf_session *session = container_of(header, struct perf_session, header);
1574         struct {
1575                 struct perf_event_header   header;
1576                 u8                         build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1577                 char                       filename[0];
1578         } old_bev;
1579         struct build_id_event bev;
1580         char filename[PATH_MAX];
1581         u64 limit = offset + size;
1582
1583         while (offset < limit) {
1584                 ssize_t len;
1585
1586                 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1587                         return -1;
1588
1589                 if (header->needs_swap)
1590                         perf_event_header__bswap(&old_bev.header);
1591
1592                 len = old_bev.header.size - sizeof(old_bev);
1593                 if (readn(input, filename, len) != len)
1594                         return -1;
1595
1596                 bev.header = old_bev.header;
1597
1598                 /*
1599                  * As the pid is the missing value, we need to fill
1600                  * it properly. The header.misc value give us nice hint.
1601                  */
1602                 bev.pid = HOST_KERNEL_ID;
1603                 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1604                     bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1605                         bev.pid = DEFAULT_GUEST_KERNEL_ID;
1606
1607                 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1608                 __event_process_build_id(&bev, filename, session);
1609
1610                 offset += bev.header.size;
1611         }
1612
1613         return 0;
1614 }
1615
1616 static int perf_header__read_build_ids(struct perf_header *header,
1617                                        int input, u64 offset, u64 size)
1618 {
1619         struct perf_session *session = container_of(header, struct perf_session, header);
1620         struct build_id_event bev;
1621         char filename[PATH_MAX];
1622         u64 limit = offset + size, orig_offset = offset;
1623         int err = -1;
1624
1625         while (offset < limit) {
1626                 ssize_t len;
1627
1628                 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1629                         goto out;
1630
1631                 if (header->needs_swap)
1632                         perf_event_header__bswap(&bev.header);
1633
1634                 len = bev.header.size - sizeof(bev);
1635                 if (readn(input, filename, len) != len)
1636                         goto out;
1637                 /*
1638                  * The a1645ce1 changeset:
1639                  *
1640                  * "perf: 'perf kvm' tool for monitoring guest performance from host"
1641                  *
1642                  * Added a field to struct build_id_event that broke the file
1643                  * format.
1644                  *
1645                  * Since the kernel build-id is the first entry, process the
1646                  * table using the old format if the well known
1647                  * '[kernel.kallsyms]' string for the kernel build-id has the
1648                  * first 4 characters chopped off (where the pid_t sits).
1649                  */
1650                 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1651                         if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1652                                 return -1;
1653                         return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1654                 }
1655
1656                 __event_process_build_id(&bev, filename, session);
1657
1658                 offset += bev.header.size;
1659         }
1660         err = 0;
1661 out:
1662         return err;
1663 }
1664
1665 /* Macro for features that simply need to read and store a string. */
1666 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \
1667 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \
1668 {\
1669         ff->ph->env.__feat_env = do_read_string(ff); \
1670         return ff->ph->env.__feat_env ? 0 : -ENOMEM; \
1671 }
1672
1673 FEAT_PROCESS_STR_FUN(hostname, hostname);
1674 FEAT_PROCESS_STR_FUN(osrelease, os_release);
1675 FEAT_PROCESS_STR_FUN(version, version);
1676 FEAT_PROCESS_STR_FUN(arch, arch);
1677 FEAT_PROCESS_STR_FUN(cpudesc, cpu_desc);
1678 FEAT_PROCESS_STR_FUN(cpuid, cpuid);
1679
1680 static int process_tracing_data(struct feat_fd *ff, void *data)
1681 {
1682         ssize_t ret = trace_report(ff->fd, data, false);
1683
1684         return ret < 0 ? -1 : 0;
1685 }
1686
1687 static int process_build_id(struct feat_fd *ff, void *data __maybe_unused)
1688 {
1689         if (perf_header__read_build_ids(ff->ph, ff->fd, ff->offset, ff->size))
1690                 pr_debug("Failed to read buildids, continuing...\n");
1691         return 0;
1692 }
1693
1694 static int process_nrcpus(struct feat_fd *ff, void *data __maybe_unused)
1695 {
1696         int ret;
1697         u32 nr_cpus_avail, nr_cpus_online;
1698
1699         ret = do_read_u32(ff, &nr_cpus_avail);
1700         if (ret)
1701                 return ret;
1702
1703         ret = do_read_u32(ff, &nr_cpus_online);
1704         if (ret)
1705                 return ret;
1706         ff->ph->env.nr_cpus_avail = (int)nr_cpus_avail;
1707         ff->ph->env.nr_cpus_online = (int)nr_cpus_online;
1708         return 0;
1709 }
1710
1711 static int process_total_mem(struct feat_fd *ff, void *data __maybe_unused)
1712 {
1713         u64 total_mem;
1714         int ret;
1715
1716         ret = do_read_u64(ff, &total_mem);
1717         if (ret)
1718                 return -1;
1719         ff->ph->env.total_mem = (unsigned long long)total_mem;
1720         return 0;
1721 }
1722
1723 static struct perf_evsel *
1724 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1725 {
1726         struct perf_evsel *evsel;
1727
1728         evlist__for_each_entry(evlist, evsel) {
1729                 if (evsel->idx == idx)
1730                         return evsel;
1731         }
1732
1733         return NULL;
1734 }
1735
1736 static void
1737 perf_evlist__set_event_name(struct perf_evlist *evlist,
1738                             struct perf_evsel *event)
1739 {
1740         struct perf_evsel *evsel;
1741
1742         if (!event->name)
1743                 return;
1744
1745         evsel = perf_evlist__find_by_index(evlist, event->idx);
1746         if (!evsel)
1747                 return;
1748
1749         if (evsel->name)
1750                 return;
1751
1752         evsel->name = strdup(event->name);
1753 }
1754
1755 static int
1756 process_event_desc(struct feat_fd *ff, void *data __maybe_unused)
1757 {
1758         struct perf_session *session;
1759         struct perf_evsel *evsel, *events = read_event_desc(ff);
1760
1761         if (!events)
1762                 return 0;
1763
1764         session = container_of(ff->ph, struct perf_session, header);
1765
1766         if (session->file->is_pipe) {
1767                 /* Save events for reading later by print_event_desc,
1768                  * since they can't be read again in pipe mode. */
1769                 ff->events = events;
1770         }
1771
1772         for (evsel = events; evsel->attr.size; evsel++)
1773                 perf_evlist__set_event_name(session->evlist, evsel);
1774
1775         if (!session->file->is_pipe)
1776                 free_event_desc(events);
1777
1778         return 0;
1779 }
1780
1781 static int process_cmdline(struct feat_fd *ff, void *data __maybe_unused)
1782 {
1783         char *str, *cmdline = NULL, **argv = NULL;
1784         u32 nr, i, len = 0;
1785
1786         if (do_read_u32(ff, &nr))
1787                 return -1;
1788
1789         ff->ph->env.nr_cmdline = nr;
1790
1791         cmdline = zalloc(ff->size + nr + 1);
1792         if (!cmdline)
1793                 return -1;
1794
1795         argv = zalloc(sizeof(char *) * (nr + 1));
1796         if (!argv)
1797                 goto error;
1798
1799         for (i = 0; i < nr; i++) {
1800                 str = do_read_string(ff);
1801                 if (!str)
1802                         goto error;
1803
1804                 argv[i] = cmdline + len;
1805                 memcpy(argv[i], str, strlen(str) + 1);
1806                 len += strlen(str) + 1;
1807                 free(str);
1808         }
1809         ff->ph->env.cmdline = cmdline;
1810         ff->ph->env.cmdline_argv = (const char **) argv;
1811         return 0;
1812
1813 error:
1814         free(argv);
1815         free(cmdline);
1816         return -1;
1817 }
1818
1819 static int process_cpu_topology(struct feat_fd *ff, void *data __maybe_unused)
1820 {
1821         u32 nr, i;
1822         char *str;
1823         struct strbuf sb;
1824         int cpu_nr = ff->ph->env.nr_cpus_avail;
1825         u64 size = 0;
1826         struct perf_header *ph = ff->ph;
1827         bool do_core_id_test = true;
1828
1829         ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
1830         if (!ph->env.cpu)
1831                 return -1;
1832
1833         if (do_read_u32(ff, &nr))
1834                 goto free_cpu;
1835
1836         ph->env.nr_sibling_cores = nr;
1837         size += sizeof(u32);
1838         if (strbuf_init(&sb, 128) < 0)
1839                 goto free_cpu;
1840
1841         for (i = 0; i < nr; i++) {
1842                 str = do_read_string(ff);
1843                 if (!str)
1844                         goto error;
1845
1846                 /* include a NULL character at the end */
1847                 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
1848                         goto error;
1849                 size += string_size(str);
1850                 free(str);
1851         }
1852         ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1853
1854         if (do_read_u32(ff, &nr))
1855                 return -1;
1856
1857         ph->env.nr_sibling_threads = nr;
1858         size += sizeof(u32);
1859
1860         for (i = 0; i < nr; i++) {
1861                 str = do_read_string(ff);
1862                 if (!str)
1863                         goto error;
1864
1865                 /* include a NULL character at the end */
1866                 if (strbuf_add(&sb, str, strlen(str) + 1) < 0)
1867                         goto error;
1868                 size += string_size(str);
1869                 free(str);
1870         }
1871         ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1872
1873         /*
1874          * The header may be from old perf,
1875          * which doesn't include core id and socket id information.
1876          */
1877         if (ff->size <= size) {
1878                 zfree(&ph->env.cpu);
1879                 return 0;
1880         }
1881
1882         /* On s390 the socket_id number is not related to the numbers of cpus.
1883          * The socket_id number might be higher than the numbers of cpus.
1884          * This depends on the configuration.
1885          */
1886         if (ph->env.arch && !strncmp(ph->env.arch, "s390", 4))
1887                 do_core_id_test = false;
1888
1889         for (i = 0; i < (u32)cpu_nr; i++) {
1890                 if (do_read_u32(ff, &nr))
1891                         goto free_cpu;
1892
1893                 ph->env.cpu[i].core_id = nr;
1894
1895                 if (do_read_u32(ff, &nr))
1896                         goto free_cpu;
1897
1898                 if (do_core_id_test && nr != (u32)-1 && nr > (u32)cpu_nr) {
1899                         pr_debug("socket_id number is too big."
1900                                  "You may need to upgrade the perf tool.\n");
1901                         goto free_cpu;
1902                 }
1903
1904                 ph->env.cpu[i].socket_id = nr;
1905         }
1906
1907         return 0;
1908
1909 error:
1910         strbuf_release(&sb);
1911 free_cpu:
1912         zfree(&ph->env.cpu);
1913         return -1;
1914 }
1915
1916 static int process_numa_topology(struct feat_fd *ff, void *data __maybe_unused)
1917 {
1918         struct numa_node *nodes, *n;
1919         u32 nr, i;
1920         char *str;
1921
1922         /* nr nodes */
1923         if (do_read_u32(ff, &nr))
1924                 return -1;
1925
1926         nodes = zalloc(sizeof(*nodes) * nr);
1927         if (!nodes)
1928                 return -ENOMEM;
1929
1930         for (i = 0; i < nr; i++) {
1931                 n = &nodes[i];
1932
1933                 /* node number */
1934                 if (do_read_u32(ff, &n->node))
1935                         goto error;
1936
1937                 if (do_read_u64(ff, &n->mem_total))
1938                         goto error;
1939
1940                 if (do_read_u64(ff, &n->mem_free))
1941                         goto error;
1942
1943                 str = do_read_string(ff);
1944                 if (!str)
1945                         goto error;
1946
1947                 n->map = cpu_map__new(str);
1948                 if (!n->map)
1949                         goto error;
1950
1951                 free(str);
1952         }
1953         ff->ph->env.nr_numa_nodes = nr;
1954         ff->ph->env.numa_nodes = nodes;
1955         return 0;
1956
1957 error:
1958         free(nodes);
1959         return -1;
1960 }
1961
1962 static int process_pmu_mappings(struct feat_fd *ff, void *data __maybe_unused)
1963 {
1964         char *name;
1965         u32 pmu_num;
1966         u32 type;
1967         struct strbuf sb;
1968
1969         if (do_read_u32(ff, &pmu_num))
1970                 return -1;
1971
1972         if (!pmu_num) {
1973                 pr_debug("pmu mappings not available\n");
1974                 return 0;
1975         }
1976
1977         ff->ph->env.nr_pmu_mappings = pmu_num;
1978         if (strbuf_init(&sb, 128) < 0)
1979                 return -1;
1980
1981         while (pmu_num) {
1982                 if (do_read_u32(ff, &type))
1983                         goto error;
1984
1985                 name = do_read_string(ff);
1986                 if (!name)
1987                         goto error;
1988
1989                 if (strbuf_addf(&sb, "%u:%s", type, name) < 0)
1990                         goto error;
1991                 /* include a NULL character at the end */
1992                 if (strbuf_add(&sb, "", 1) < 0)
1993                         goto error;
1994
1995                 if (!strcmp(name, "msr"))
1996                         ff->ph->env.msr_pmu_type = type;
1997
1998                 free(name);
1999                 pmu_num--;
2000         }
2001         ff->ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
2002         return 0;
2003
2004 error:
2005         strbuf_release(&sb);
2006         return -1;
2007 }
2008
2009 static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused)
2010 {
2011         size_t ret = -1;
2012         u32 i, nr, nr_groups;
2013         struct perf_session *session;
2014         struct perf_evsel *evsel, *leader = NULL;
2015         struct group_desc {
2016                 char *name;
2017                 u32 leader_idx;
2018                 u32 nr_members;
2019         } *desc;
2020
2021         if (do_read_u32(ff, &nr_groups))
2022                 return -1;
2023
2024         ff->ph->env.nr_groups = nr_groups;
2025         if (!nr_groups) {
2026                 pr_debug("group desc not available\n");
2027                 return 0;
2028         }
2029
2030         desc = calloc(nr_groups, sizeof(*desc));
2031         if (!desc)
2032                 return -1;
2033
2034         for (i = 0; i < nr_groups; i++) {
2035                 desc[i].name = do_read_string(ff);
2036                 if (!desc[i].name)
2037                         goto out_free;
2038
2039                 if (do_read_u32(ff, &desc[i].leader_idx))
2040                         goto out_free;
2041
2042                 if (do_read_u32(ff, &desc[i].nr_members))
2043                         goto out_free;
2044         }
2045
2046         /*
2047          * Rebuild group relationship based on the group_desc
2048          */
2049         session = container_of(ff->ph, struct perf_session, header);
2050         session->evlist->nr_groups = nr_groups;
2051
2052         i = nr = 0;
2053         evlist__for_each_entry(session->evlist, evsel) {
2054                 if (evsel->idx == (int) desc[i].leader_idx) {
2055                         evsel->leader = evsel;
2056                         /* {anon_group} is a dummy name */
2057                         if (strcmp(desc[i].name, "{anon_group}")) {
2058                                 evsel->group_name = desc[i].name;
2059                                 desc[i].name = NULL;
2060                         }
2061                         evsel->nr_members = desc[i].nr_members;
2062
2063                         if (i >= nr_groups || nr > 0) {
2064                                 pr_debug("invalid group desc\n");
2065                                 goto out_free;
2066                         }
2067
2068                         leader = evsel;
2069                         nr = evsel->nr_members - 1;
2070                         i++;
2071                 } else if (nr) {
2072                         /* This is a group member */
2073                         evsel->leader = leader;
2074
2075                         nr--;
2076                 }
2077         }
2078
2079         if (i != nr_groups || nr != 0) {
2080                 pr_debug("invalid group desc\n");
2081                 goto out_free;
2082         }
2083
2084         ret = 0;
2085 out_free:
2086         for (i = 0; i < nr_groups; i++)
2087                 zfree(&desc[i].name);
2088         free(desc);
2089
2090         return ret;
2091 }
2092
2093 static int process_auxtrace(struct feat_fd *ff, void *data __maybe_unused)
2094 {
2095         struct perf_session *session;
2096         int err;
2097
2098         session = container_of(ff->ph, struct perf_session, header);
2099
2100         err = auxtrace_index__process(ff->fd, ff->size, session,
2101                                       ff->ph->needs_swap);
2102         if (err < 0)
2103                 pr_err("Failed to process auxtrace index\n");
2104         return err;
2105 }
2106
2107 static int process_cache(struct feat_fd *ff, void *data __maybe_unused)
2108 {
2109         struct cpu_cache_level *caches;
2110         u32 cnt, i, version;
2111
2112         if (do_read_u32(ff, &version))
2113                 return -1;
2114
2115         if (version != 1)
2116                 return -1;
2117
2118         if (do_read_u32(ff, &cnt))
2119                 return -1;
2120
2121         caches = zalloc(sizeof(*caches) * cnt);
2122         if (!caches)
2123                 return -1;
2124
2125         for (i = 0; i < cnt; i++) {
2126                 struct cpu_cache_level c;
2127
2128                 #define _R(v)                                           \
2129                         if (do_read_u32(ff, &c.v))\
2130                                 goto out_free_caches;                   \
2131
2132                 _R(level)
2133                 _R(line_size)
2134                 _R(sets)
2135                 _R(ways)
2136                 #undef _R
2137
2138                 #define _R(v)                                   \
2139                         c.v = do_read_string(ff);               \
2140                         if (!c.v)                               \
2141                                 goto out_free_caches;
2142
2143                 _R(type)
2144                 _R(size)
2145                 _R(map)
2146                 #undef _R
2147
2148                 caches[i] = c;
2149         }
2150
2151         ff->ph->env.caches = caches;
2152         ff->ph->env.caches_cnt = cnt;
2153         return 0;
2154 out_free_caches:
2155         free(caches);
2156         return -1;
2157 }
2158
2159 struct feature_ops {
2160         int (*write)(struct feat_fd *ff, struct perf_evlist *evlist);
2161         void (*print)(struct feat_fd *ff, FILE *fp);
2162         int (*process)(struct feat_fd *ff, void *data);
2163         const char *name;
2164         bool full_only;
2165         bool synthesize;
2166 };
2167
2168 #define FEAT_OPR(n, func, __full_only) \
2169         [HEADER_##n] = {                                        \
2170                 .name       = __stringify(n),                   \
2171                 .write      = write_##func,                     \
2172                 .print      = print_##func,                     \
2173                 .full_only  = __full_only,                      \
2174                 .process    = process_##func,                   \
2175                 .synthesize = true                              \
2176         }
2177
2178 #define FEAT_OPN(n, func, __full_only) \
2179         [HEADER_##n] = {                                        \
2180                 .name       = __stringify(n),                   \
2181                 .write      = write_##func,                     \
2182                 .print      = print_##func,                     \
2183                 .full_only  = __full_only,                      \
2184                 .process    = process_##func                    \
2185         }
2186
2187 /* feature_ops not implemented: */
2188 #define print_tracing_data      NULL
2189 #define print_build_id          NULL
2190
2191 #define process_branch_stack    NULL
2192 #define process_stat            NULL
2193
2194
2195 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
2196         FEAT_OPN(TRACING_DATA,  tracing_data,   false),
2197         FEAT_OPN(BUILD_ID,      build_id,       false),
2198         FEAT_OPR(HOSTNAME,      hostname,       false),
2199         FEAT_OPR(OSRELEASE,     osrelease,      false),
2200         FEAT_OPR(VERSION,       version,        false),
2201         FEAT_OPR(ARCH,          arch,           false),
2202         FEAT_OPR(NRCPUS,        nrcpus,         false),
2203         FEAT_OPR(CPUDESC,       cpudesc,        false),
2204         FEAT_OPR(CPUID,         cpuid,          false),
2205         FEAT_OPR(TOTAL_MEM,     total_mem,      false),
2206         FEAT_OPR(EVENT_DESC,    event_desc,     false),
2207         FEAT_OPR(CMDLINE,       cmdline,        false),
2208         FEAT_OPR(CPU_TOPOLOGY,  cpu_topology,   true),
2209         FEAT_OPR(NUMA_TOPOLOGY, numa_topology,  true),
2210         FEAT_OPN(BRANCH_STACK,  branch_stack,   false),
2211         FEAT_OPR(PMU_MAPPINGS,  pmu_mappings,   false),
2212         FEAT_OPN(GROUP_DESC,    group_desc,     false),
2213         FEAT_OPN(AUXTRACE,      auxtrace,       false),
2214         FEAT_OPN(STAT,          stat,           false),
2215         FEAT_OPN(CACHE,         cache,          true),
2216 };
2217
2218 struct header_print_data {
2219         FILE *fp;
2220         bool full; /* extended list of headers */
2221 };
2222
2223 static int perf_file_section__fprintf_info(struct perf_file_section *section,
2224                                            struct perf_header *ph,
2225                                            int feat, int fd, void *data)
2226 {
2227         struct header_print_data *hd = data;
2228         struct feat_fd ff;
2229
2230         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2231                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2232                                 "%d, continuing...\n", section->offset, feat);
2233                 return 0;
2234         }
2235         if (feat >= HEADER_LAST_FEATURE) {
2236                 pr_warning("unknown feature %d\n", feat);
2237                 return 0;
2238         }
2239         if (!feat_ops[feat].print)
2240                 return 0;
2241
2242         ff = (struct  feat_fd) {
2243                 .fd = fd,
2244                 .ph = ph,
2245         };
2246
2247         if (!feat_ops[feat].full_only || hd->full)
2248                 feat_ops[feat].print(&ff, hd->fp);
2249         else
2250                 fprintf(hd->fp, "# %s info available, use -I to display\n",
2251                         feat_ops[feat].name);
2252
2253         return 0;
2254 }
2255
2256 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
2257 {
2258         struct header_print_data hd;
2259         struct perf_header *header = &session->header;
2260         int fd = perf_data_file__fd(session->file);
2261         struct stat st;
2262         int ret, bit;
2263
2264         hd.fp = fp;
2265         hd.full = full;
2266
2267         ret = fstat(fd, &st);
2268         if (ret == -1)
2269                 return -1;
2270
2271         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
2272
2273         perf_header__process_sections(header, fd, &hd,
2274                                       perf_file_section__fprintf_info);
2275
2276         if (session->file->is_pipe)
2277                 return 0;
2278
2279         fprintf(fp, "# missing features: ");
2280         for_each_clear_bit(bit, header->adds_features, HEADER_LAST_FEATURE) {
2281                 if (bit)
2282                         fprintf(fp, "%s ", feat_ops[bit].name);
2283         }
2284
2285         fprintf(fp, "\n");
2286         return 0;
2287 }
2288
2289 static int do_write_feat(struct feat_fd *ff, int type,
2290                          struct perf_file_section **p,
2291                          struct perf_evlist *evlist)
2292 {
2293         int err;
2294         int ret = 0;
2295
2296         if (perf_header__has_feat(ff->ph, type)) {
2297                 if (!feat_ops[type].write)
2298                         return -1;
2299
2300                 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__))
2301                         return -1;
2302
2303                 (*p)->offset = lseek(ff->fd, 0, SEEK_CUR);
2304
2305                 err = feat_ops[type].write(ff, evlist);
2306                 if (err < 0) {
2307                         pr_debug("failed to write feature %s\n", feat_ops[type].name);
2308
2309                         /* undo anything written */
2310                         lseek(ff->fd, (*p)->offset, SEEK_SET);
2311
2312                         return -1;
2313                 }
2314                 (*p)->size = lseek(ff->fd, 0, SEEK_CUR) - (*p)->offset;
2315                 (*p)++;
2316         }
2317         return ret;
2318 }
2319
2320 static int perf_header__adds_write(struct perf_header *header,
2321                                    struct perf_evlist *evlist, int fd)
2322 {
2323         int nr_sections;
2324         struct feat_fd ff;
2325         struct perf_file_section *feat_sec, *p;
2326         int sec_size;
2327         u64 sec_start;
2328         int feat;
2329         int err;
2330
2331         ff = (struct feat_fd){
2332                 .fd  = fd,
2333                 .ph = header,
2334         };
2335
2336         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2337         if (!nr_sections)
2338                 return 0;
2339
2340         feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2341         if (feat_sec == NULL)
2342                 return -ENOMEM;
2343
2344         sec_size = sizeof(*feat_sec) * nr_sections;
2345
2346         sec_start = header->feat_offset;
2347         lseek(fd, sec_start + sec_size, SEEK_SET);
2348
2349         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2350                 if (do_write_feat(&ff, feat, &p, evlist))
2351                         perf_header__clear_feat(header, feat);
2352         }
2353
2354         lseek(fd, sec_start, SEEK_SET);
2355         /*
2356          * may write more than needed due to dropped feature, but
2357          * this is okay, reader will skip the mising entries
2358          */
2359         err = do_write(&ff, feat_sec, sec_size);
2360         if (err < 0)
2361                 pr_debug("failed to write feature section\n");
2362         free(feat_sec);
2363         return err;
2364 }
2365
2366 int perf_header__write_pipe(int fd)
2367 {
2368         struct perf_pipe_file_header f_header;
2369         struct feat_fd ff;
2370         int err;
2371
2372         ff = (struct feat_fd){ .fd = fd };
2373
2374         f_header = (struct perf_pipe_file_header){
2375                 .magic     = PERF_MAGIC,
2376                 .size      = sizeof(f_header),
2377         };
2378
2379         err = do_write(&ff, &f_header, sizeof(f_header));
2380         if (err < 0) {
2381                 pr_debug("failed to write perf pipe header\n");
2382                 return err;
2383         }
2384
2385         return 0;
2386 }
2387
2388 int perf_session__write_header(struct perf_session *session,
2389                                struct perf_evlist *evlist,
2390                                int fd, bool at_exit)
2391 {
2392         struct perf_file_header f_header;
2393         struct perf_file_attr   f_attr;
2394         struct perf_header *header = &session->header;
2395         struct perf_evsel *evsel;
2396         struct feat_fd ff;
2397         u64 attr_offset;
2398         int err;
2399
2400         ff = (struct feat_fd){ .fd = fd};
2401         lseek(fd, sizeof(f_header), SEEK_SET);
2402
2403         evlist__for_each_entry(session->evlist, evsel) {
2404                 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2405                 err = do_write(&ff, evsel->id, evsel->ids * sizeof(u64));
2406                 if (err < 0) {
2407                         pr_debug("failed to write perf header\n");
2408                         return err;
2409                 }
2410         }
2411
2412         attr_offset = lseek(ff.fd, 0, SEEK_CUR);
2413
2414         evlist__for_each_entry(evlist, evsel) {
2415                 f_attr = (struct perf_file_attr){
2416                         .attr = evsel->attr,
2417                         .ids  = {
2418                                 .offset = evsel->id_offset,
2419                                 .size   = evsel->ids * sizeof(u64),
2420                         }
2421                 };
2422                 err = do_write(&ff, &f_attr, sizeof(f_attr));
2423                 if (err < 0) {
2424                         pr_debug("failed to write perf header attribute\n");
2425                         return err;
2426                 }
2427         }
2428
2429         if (!header->data_offset)
2430                 header->data_offset = lseek(fd, 0, SEEK_CUR);
2431         header->feat_offset = header->data_offset + header->data_size;
2432
2433         if (at_exit) {
2434                 err = perf_header__adds_write(header, evlist, fd);
2435                 if (err < 0)
2436                         return err;
2437         }
2438
2439         f_header = (struct perf_file_header){
2440                 .magic     = PERF_MAGIC,
2441                 .size      = sizeof(f_header),
2442                 .attr_size = sizeof(f_attr),
2443                 .attrs = {
2444                         .offset = attr_offset,
2445                         .size   = evlist->nr_entries * sizeof(f_attr),
2446                 },
2447                 .data = {
2448                         .offset = header->data_offset,
2449                         .size   = header->data_size,
2450                 },
2451                 /* event_types is ignored, store zeros */
2452         };
2453
2454         memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2455
2456         lseek(fd, 0, SEEK_SET);
2457         err = do_write(&ff, &f_header, sizeof(f_header));
2458         if (err < 0) {
2459                 pr_debug("failed to write perf header\n");
2460                 return err;
2461         }
2462         lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2463
2464         return 0;
2465 }
2466
2467 static int perf_header__getbuffer64(struct perf_header *header,
2468                                     int fd, void *buf, size_t size)
2469 {
2470         if (readn(fd, buf, size) <= 0)
2471                 return -1;
2472
2473         if (header->needs_swap)
2474                 mem_bswap_64(buf, size);
2475
2476         return 0;
2477 }
2478
2479 int perf_header__process_sections(struct perf_header *header, int fd,
2480                                   void *data,
2481                                   int (*process)(struct perf_file_section *section,
2482                                                  struct perf_header *ph,
2483                                                  int feat, int fd, void *data))
2484 {
2485         struct perf_file_section *feat_sec, *sec;
2486         int nr_sections;
2487         int sec_size;
2488         int feat;
2489         int err;
2490
2491         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2492         if (!nr_sections)
2493                 return 0;
2494
2495         feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2496         if (!feat_sec)
2497                 return -1;
2498
2499         sec_size = sizeof(*feat_sec) * nr_sections;
2500
2501         lseek(fd, header->feat_offset, SEEK_SET);
2502
2503         err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2504         if (err < 0)
2505                 goto out_free;
2506
2507         for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2508                 err = process(sec++, header, feat, fd, data);
2509                 if (err < 0)
2510                         goto out_free;
2511         }
2512         err = 0;
2513 out_free:
2514         free(feat_sec);
2515         return err;
2516 }
2517
2518 static const int attr_file_abi_sizes[] = {
2519         [0] = PERF_ATTR_SIZE_VER0,
2520         [1] = PERF_ATTR_SIZE_VER1,
2521         [2] = PERF_ATTR_SIZE_VER2,
2522         [3] = PERF_ATTR_SIZE_VER3,
2523         [4] = PERF_ATTR_SIZE_VER4,
2524         0,
2525 };
2526
2527 /*
2528  * In the legacy file format, the magic number is not used to encode endianness.
2529  * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2530  * on ABI revisions, we need to try all combinations for all endianness to
2531  * detect the endianness.
2532  */
2533 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2534 {
2535         uint64_t ref_size, attr_size;
2536         int i;
2537
2538         for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2539                 ref_size = attr_file_abi_sizes[i]
2540                          + sizeof(struct perf_file_section);
2541                 if (hdr_sz != ref_size) {
2542                         attr_size = bswap_64(hdr_sz);
2543                         if (attr_size != ref_size)
2544                                 continue;
2545
2546                         ph->needs_swap = true;
2547                 }
2548                 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2549                          i,
2550                          ph->needs_swap);
2551                 return 0;
2552         }
2553         /* could not determine endianness */
2554         return -1;
2555 }
2556
2557 #define PERF_PIPE_HDR_VER0      16
2558
2559 static const size_t attr_pipe_abi_sizes[] = {
2560         [0] = PERF_PIPE_HDR_VER0,
2561         0,
2562 };
2563
2564 /*
2565  * In the legacy pipe format, there is an implicit assumption that endiannesss
2566  * between host recording the samples, and host parsing the samples is the
2567  * same. This is not always the case given that the pipe output may always be
2568  * redirected into a file and analyzed on a different machine with possibly a
2569  * different endianness and perf_event ABI revsions in the perf tool itself.
2570  */
2571 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2572 {
2573         u64 attr_size;
2574         int i;
2575
2576         for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2577                 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2578                         attr_size = bswap_64(hdr_sz);
2579                         if (attr_size != hdr_sz)
2580                                 continue;
2581
2582                         ph->needs_swap = true;
2583                 }
2584                 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2585                 return 0;
2586         }
2587         return -1;
2588 }
2589
2590 bool is_perf_magic(u64 magic)
2591 {
2592         if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2593                 || magic == __perf_magic2
2594                 || magic == __perf_magic2_sw)
2595                 return true;
2596
2597         return false;
2598 }
2599
2600 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2601                               bool is_pipe, struct perf_header *ph)
2602 {
2603         int ret;
2604
2605         /* check for legacy format */
2606         ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2607         if (ret == 0) {
2608                 ph->version = PERF_HEADER_VERSION_1;
2609                 pr_debug("legacy perf.data format\n");
2610                 if (is_pipe)
2611                         return try_all_pipe_abis(hdr_sz, ph);
2612
2613                 return try_all_file_abis(hdr_sz, ph);
2614         }
2615         /*
2616          * the new magic number serves two purposes:
2617          * - unique number to identify actual perf.data files
2618          * - encode endianness of file
2619          */
2620         ph->version = PERF_HEADER_VERSION_2;
2621
2622         /* check magic number with one endianness */
2623         if (magic == __perf_magic2)
2624                 return 0;
2625
2626         /* check magic number with opposite endianness */
2627         if (magic != __perf_magic2_sw)
2628                 return -1;
2629
2630         ph->needs_swap = true;
2631
2632         return 0;
2633 }
2634
2635 int perf_file_header__read(struct perf_file_header *header,
2636                            struct perf_header *ph, int fd)
2637 {
2638         ssize_t ret;
2639
2640         lseek(fd, 0, SEEK_SET);
2641
2642         ret = readn(fd, header, sizeof(*header));
2643         if (ret <= 0)
2644                 return -1;
2645
2646         if (check_magic_endian(header->magic,
2647                                header->attr_size, false, ph) < 0) {
2648                 pr_debug("magic/endian check failed\n");
2649                 return -1;
2650         }
2651
2652         if (ph->needs_swap) {
2653                 mem_bswap_64(header, offsetof(struct perf_file_header,
2654                              adds_features));
2655         }
2656
2657         if (header->size != sizeof(*header)) {
2658                 /* Support the previous format */
2659                 if (header->size == offsetof(typeof(*header), adds_features))
2660                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2661                 else
2662                         return -1;
2663         } else if (ph->needs_swap) {
2664                 /*
2665                  * feature bitmap is declared as an array of unsigned longs --
2666                  * not good since its size can differ between the host that
2667                  * generated the data file and the host analyzing the file.
2668                  *
2669                  * We need to handle endianness, but we don't know the size of
2670                  * the unsigned long where the file was generated. Take a best
2671                  * guess at determining it: try 64-bit swap first (ie., file
2672                  * created on a 64-bit host), and check if the hostname feature
2673                  * bit is set (this feature bit is forced on as of fbe96f2).
2674                  * If the bit is not, undo the 64-bit swap and try a 32-bit
2675                  * swap. If the hostname bit is still not set (e.g., older data
2676                  * file), punt and fallback to the original behavior --
2677                  * clearing all feature bits and setting buildid.
2678                  */
2679                 mem_bswap_64(&header->adds_features,
2680                             BITS_TO_U64(HEADER_FEAT_BITS));
2681
2682                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2683                         /* unswap as u64 */
2684                         mem_bswap_64(&header->adds_features,
2685                                     BITS_TO_U64(HEADER_FEAT_BITS));
2686
2687                         /* unswap as u32 */
2688                         mem_bswap_32(&header->adds_features,
2689                                     BITS_TO_U32(HEADER_FEAT_BITS));
2690                 }
2691
2692                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2693                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2694                         set_bit(HEADER_BUILD_ID, header->adds_features);
2695                 }
2696         }
2697
2698         memcpy(&ph->adds_features, &header->adds_features,
2699                sizeof(ph->adds_features));
2700
2701         ph->data_offset  = header->data.offset;
2702         ph->data_size    = header->data.size;
2703         ph->feat_offset  = header->data.offset + header->data.size;
2704         return 0;
2705 }
2706
2707 static int perf_file_section__process(struct perf_file_section *section,
2708                                       struct perf_header *ph,
2709                                       int feat, int fd, void *data)
2710 {
2711         struct feat_fd fdd = {
2712                 .fd     = fd,
2713                 .ph     = ph,
2714                 .size   = section->size,
2715                 .offset = section->offset,
2716         };
2717
2718         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2719                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2720                           "%d, continuing...\n", section->offset, feat);
2721                 return 0;
2722         }
2723
2724         if (feat >= HEADER_LAST_FEATURE) {
2725                 pr_debug("unknown feature %d, continuing...\n", feat);
2726                 return 0;
2727         }
2728
2729         if (!feat_ops[feat].process)
2730                 return 0;
2731
2732         return feat_ops[feat].process(&fdd, data);
2733 }
2734
2735 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2736                                        struct perf_header *ph, int fd,
2737                                        bool repipe)
2738 {
2739         struct feat_fd ff = {
2740                 .fd = STDOUT_FILENO,
2741                 .ph = ph,
2742         };
2743         ssize_t ret;
2744
2745         ret = readn(fd, header, sizeof(*header));
2746         if (ret <= 0)
2747                 return -1;
2748
2749         if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2750                 pr_debug("endian/magic failed\n");
2751                 return -1;
2752         }
2753
2754         if (ph->needs_swap)
2755                 header->size = bswap_64(header->size);
2756
2757         if (repipe && do_write(&ff, header, sizeof(*header)) < 0)
2758                 return -1;
2759
2760         return 0;
2761 }
2762
2763 static int perf_header__read_pipe(struct perf_session *session)
2764 {
2765         struct perf_header *header = &session->header;
2766         struct perf_pipe_file_header f_header;
2767
2768         if (perf_file_header__read_pipe(&f_header, header,
2769                                         perf_data_file__fd(session->file),
2770                                         session->repipe) < 0) {
2771                 pr_debug("incompatible file format\n");
2772                 return -EINVAL;
2773         }
2774
2775         return 0;
2776 }
2777
2778 static int read_attr(int fd, struct perf_header *ph,
2779                      struct perf_file_attr *f_attr)
2780 {
2781         struct perf_event_attr *attr = &f_attr->attr;
2782         size_t sz, left;
2783         size_t our_sz = sizeof(f_attr->attr);
2784         ssize_t ret;
2785
2786         memset(f_attr, 0, sizeof(*f_attr));
2787
2788         /* read minimal guaranteed structure */
2789         ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2790         if (ret <= 0) {
2791                 pr_debug("cannot read %d bytes of header attr\n",
2792                          PERF_ATTR_SIZE_VER0);
2793                 return -1;
2794         }
2795
2796         /* on file perf_event_attr size */
2797         sz = attr->size;
2798
2799         if (ph->needs_swap)
2800                 sz = bswap_32(sz);
2801
2802         if (sz == 0) {
2803                 /* assume ABI0 */
2804                 sz =  PERF_ATTR_SIZE_VER0;
2805         } else if (sz > our_sz) {
2806                 pr_debug("file uses a more recent and unsupported ABI"
2807                          " (%zu bytes extra)\n", sz - our_sz);
2808                 return -1;
2809         }
2810         /* what we have not yet read and that we know about */
2811         left = sz - PERF_ATTR_SIZE_VER0;
2812         if (left) {
2813                 void *ptr = attr;
2814                 ptr += PERF_ATTR_SIZE_VER0;
2815
2816                 ret = readn(fd, ptr, left);
2817         }
2818         /* read perf_file_section, ids are read in caller */
2819         ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2820
2821         return ret <= 0 ? -1 : 0;
2822 }
2823
2824 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2825                                                 struct pevent *pevent)
2826 {
2827         struct event_format *event;
2828         char bf[128];
2829
2830         /* already prepared */
2831         if (evsel->tp_format)
2832                 return 0;
2833
2834         if (pevent == NULL) {
2835                 pr_debug("broken or missing trace data\n");
2836                 return -1;
2837         }
2838
2839         event = pevent_find_event(pevent, evsel->attr.config);
2840         if (event == NULL) {
2841                 pr_debug("cannot find event format for %d\n", (int)evsel->attr.config);
2842                 return -1;
2843         }
2844
2845         if (!evsel->name) {
2846                 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2847                 evsel->name = strdup(bf);
2848                 if (evsel->name == NULL)
2849                         return -1;
2850         }
2851
2852         evsel->tp_format = event;
2853         return 0;
2854 }
2855
2856 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2857                                                   struct pevent *pevent)
2858 {
2859         struct perf_evsel *pos;
2860
2861         evlist__for_each_entry(evlist, pos) {
2862                 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2863                     perf_evsel__prepare_tracepoint_event(pos, pevent))
2864                         return -1;
2865         }
2866
2867         return 0;
2868 }
2869
2870 int perf_session__read_header(struct perf_session *session)
2871 {
2872         struct perf_data_file *file = session->file;
2873         struct perf_header *header = &session->header;
2874         struct perf_file_header f_header;
2875         struct perf_file_attr   f_attr;
2876         u64                     f_id;
2877         int nr_attrs, nr_ids, i, j;
2878         int fd = perf_data_file__fd(file);
2879
2880         session->evlist = perf_evlist__new();
2881         if (session->evlist == NULL)
2882                 return -ENOMEM;
2883
2884         session->evlist->env = &header->env;
2885         session->machines.host.env = &header->env;
2886         if (perf_data_file__is_pipe(file))
2887                 return perf_header__read_pipe(session);
2888
2889         if (perf_file_header__read(&f_header, header, fd) < 0)
2890                 return -EINVAL;
2891
2892         /*
2893          * Sanity check that perf.data was written cleanly; data size is
2894          * initialized to 0 and updated only if the on_exit function is run.
2895          * If data size is still 0 then the file contains only partial
2896          * information.  Just warn user and process it as much as it can.
2897          */
2898         if (f_header.data.size == 0) {
2899                 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2900                            "Was the 'perf record' command properly terminated?\n",
2901                            file->path);
2902         }
2903
2904         nr_attrs = f_header.attrs.size / f_header.attr_size;
2905         lseek(fd, f_header.attrs.offset, SEEK_SET);
2906
2907         for (i = 0; i < nr_attrs; i++) {
2908                 struct perf_evsel *evsel;
2909                 off_t tmp;
2910
2911                 if (read_attr(fd, header, &f_attr) < 0)
2912                         goto out_errno;
2913
2914                 if (header->needs_swap) {
2915                         f_attr.ids.size   = bswap_64(f_attr.ids.size);
2916                         f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2917                         perf_event__attr_swap(&f_attr.attr);
2918                 }
2919
2920                 tmp = lseek(fd, 0, SEEK_CUR);
2921                 evsel = perf_evsel__new(&f_attr.attr);
2922
2923                 if (evsel == NULL)
2924                         goto out_delete_evlist;
2925
2926                 evsel->needs_swap = header->needs_swap;
2927                 /*
2928                  * Do it before so that if perf_evsel__alloc_id fails, this
2929                  * entry gets purged too at perf_evlist__delete().
2930                  */
2931                 perf_evlist__add(session->evlist, evsel);
2932
2933                 nr_ids = f_attr.ids.size / sizeof(u64);
2934                 /*
2935                  * We don't have the cpu and thread maps on the header, so
2936                  * for allocating the perf_sample_id table we fake 1 cpu and
2937                  * hattr->ids threads.
2938                  */
2939                 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2940                         goto out_delete_evlist;
2941
2942                 lseek(fd, f_attr.ids.offset, SEEK_SET);
2943
2944                 for (j = 0; j < nr_ids; j++) {
2945                         if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2946                                 goto out_errno;
2947
2948                         perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2949                 }
2950
2951                 lseek(fd, tmp, SEEK_SET);
2952         }
2953
2954         symbol_conf.nr_events = nr_attrs;
2955
2956         perf_header__process_sections(header, fd, &session->tevent,
2957                                       perf_file_section__process);
2958
2959         if (perf_evlist__prepare_tracepoint_events(session->evlist,
2960                                                    session->tevent.pevent))
2961                 goto out_delete_evlist;
2962
2963         return 0;
2964 out_errno:
2965         return -errno;
2966
2967 out_delete_evlist:
2968         perf_evlist__delete(session->evlist);
2969         session->evlist = NULL;
2970         return -ENOMEM;
2971 }
2972
2973 int perf_event__synthesize_attr(struct perf_tool *tool,
2974                                 struct perf_event_attr *attr, u32 ids, u64 *id,
2975                                 perf_event__handler_t process)
2976 {
2977         union perf_event *ev;
2978         size_t size;
2979         int err;
2980
2981         size = sizeof(struct perf_event_attr);
2982         size = PERF_ALIGN(size, sizeof(u64));
2983         size += sizeof(struct perf_event_header);
2984         size += ids * sizeof(u64);
2985
2986         ev = malloc(size);
2987
2988         if (ev == NULL)
2989                 return -ENOMEM;
2990
2991         ev->attr.attr = *attr;
2992         memcpy(ev->attr.id, id, ids * sizeof(u64));
2993
2994         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2995         ev->attr.header.size = (u16)size;
2996
2997         if (ev->attr.header.size == size)
2998                 err = process(tool, ev, NULL, NULL);
2999         else
3000                 err = -E2BIG;
3001
3002         free(ev);
3003
3004         return err;
3005 }
3006
3007 int perf_event__synthesize_features(struct perf_tool *tool,
3008                                     struct perf_session *session,
3009                                     struct perf_evlist *evlist,
3010                                     perf_event__handler_t process)
3011 {
3012         struct perf_header *header = &session->header;
3013         struct feat_fd ff;
3014         struct feature_event *fe;
3015         size_t sz, sz_hdr;
3016         int feat, ret;
3017
3018         sz_hdr = sizeof(fe->header);
3019         sz = sizeof(union perf_event);
3020         /* get a nice alignment */
3021         sz = PERF_ALIGN(sz, page_size);
3022
3023         memset(&ff, 0, sizeof(ff));
3024
3025         ff.buf = malloc(sz);
3026         if (!ff.buf)
3027                 return -ENOMEM;
3028
3029         ff.size = sz - sz_hdr;
3030
3031         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
3032                 if (!feat_ops[feat].synthesize) {
3033                         pr_debug("No record header feature for header :%d\n", feat);
3034                         continue;
3035                 }
3036
3037                 ff.offset = sizeof(*fe);
3038
3039                 ret = feat_ops[feat].write(&ff, evlist);
3040                 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) {
3041                         pr_debug("Error writing feature\n");
3042                         continue;
3043                 }
3044                 /* ff.buf may have changed due to realloc in do_write() */
3045                 fe = ff.buf;
3046                 memset(fe, 0, sizeof(*fe));
3047
3048                 fe->feat_id = feat;
3049                 fe->header.type = PERF_RECORD_HEADER_FEATURE;
3050                 fe->header.size = ff.offset;
3051
3052                 ret = process(tool, ff.buf, NULL, NULL);
3053                 if (ret) {
3054                         free(ff.buf);
3055                         return ret;
3056                 }
3057         }
3058         free(ff.buf);
3059         return 0;
3060 }
3061
3062 int perf_event__process_feature(struct perf_tool *tool,
3063                                 union perf_event *event,
3064                                 struct perf_session *session __maybe_unused)
3065 {
3066         struct feat_fd ff = { .fd = 0 };
3067         struct feature_event *fe = (struct feature_event *)event;
3068         int type = fe->header.type;
3069         u64 feat = fe->feat_id;
3070
3071         if (type < 0 || type >= PERF_RECORD_HEADER_MAX) {
3072                 pr_warning("invalid record type %d in pipe-mode\n", type);
3073                 return 0;
3074         }
3075         if (feat == HEADER_RESERVED || feat > HEADER_LAST_FEATURE) {
3076                 pr_warning("invalid record type %d in pipe-mode\n", type);
3077                 return -1;
3078         }
3079
3080         if (!feat_ops[feat].process)
3081                 return 0;
3082
3083         ff.buf  = (void *)fe->data;
3084         ff.size = event->header.size - sizeof(event->header);
3085         ff.ph = &session->header;
3086
3087         if (feat_ops[feat].process(&ff, NULL))
3088                 return -1;
3089
3090         if (!feat_ops[feat].print || !tool->show_feat_hdr)
3091                 return 0;
3092
3093         if (!feat_ops[feat].full_only ||
3094             tool->show_feat_hdr >= SHOW_FEAT_HEADER_FULL_INFO) {
3095                 feat_ops[feat].print(&ff, stdout);
3096         } else {
3097                 fprintf(stdout, "# %s info available, use -I to display\n",
3098                         feat_ops[feat].name);
3099         }
3100
3101         return 0;
3102 }
3103
3104 static struct event_update_event *
3105 event_update_event__new(size_t size, u64 type, u64 id)
3106 {
3107         struct event_update_event *ev;
3108
3109         size += sizeof(*ev);
3110         size  = PERF_ALIGN(size, sizeof(u64));
3111
3112         ev = zalloc(size);
3113         if (ev) {
3114                 ev->header.type = PERF_RECORD_EVENT_UPDATE;
3115                 ev->header.size = (u16)size;
3116                 ev->type = type;
3117                 ev->id = id;
3118         }
3119         return ev;
3120 }
3121
3122 int
3123 perf_event__synthesize_event_update_unit(struct perf_tool *tool,
3124                                          struct perf_evsel *evsel,
3125                                          perf_event__handler_t process)
3126 {
3127         struct event_update_event *ev;
3128         size_t size = strlen(evsel->unit);
3129         int err;
3130
3131         ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->id[0]);
3132         if (ev == NULL)
3133                 return -ENOMEM;
3134
3135         strncpy(ev->data, evsel->unit, size);
3136         err = process(tool, (union perf_event *)ev, NULL, NULL);
3137         free(ev);
3138         return err;
3139 }
3140
3141 int
3142 perf_event__synthesize_event_update_scale(struct perf_tool *tool,
3143                                           struct perf_evsel *evsel,
3144                                           perf_event__handler_t process)
3145 {
3146         struct event_update_event *ev;
3147         struct event_update_event_scale *ev_data;
3148         int err;
3149
3150         ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->id[0]);
3151         if (ev == NULL)
3152                 return -ENOMEM;
3153
3154         ev_data = (struct event_update_event_scale *) ev->data;
3155         ev_data->scale = evsel->scale;
3156         err = process(tool, (union perf_event*) ev, NULL, NULL);
3157         free(ev);
3158         return err;
3159 }
3160
3161 int
3162 perf_event__synthesize_event_update_name(struct perf_tool *tool,
3163                                          struct perf_evsel *evsel,
3164                                          perf_event__handler_t process)
3165 {
3166         struct event_update_event *ev;
3167         size_t len = strlen(evsel->name);
3168         int err;
3169
3170         ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->id[0]);
3171         if (ev == NULL)
3172                 return -ENOMEM;
3173
3174         strncpy(ev->data, evsel->name, len);
3175         err = process(tool, (union perf_event*) ev, NULL, NULL);
3176         free(ev);
3177         return err;
3178 }
3179
3180 int
3181 perf_event__synthesize_event_update_cpus(struct perf_tool *tool,
3182                                         struct perf_evsel *evsel,
3183                                         perf_event__handler_t process)
3184 {
3185         size_t size = sizeof(struct event_update_event);
3186         struct event_update_event *ev;
3187         int max, err;
3188         u16 type;
3189
3190         if (!evsel->own_cpus)
3191                 return 0;
3192
3193         ev = cpu_map_data__alloc(evsel->own_cpus, &size, &type, &max);
3194         if (!ev)
3195                 return -ENOMEM;
3196
3197         ev->header.type = PERF_RECORD_EVENT_UPDATE;
3198         ev->header.size = (u16)size;
3199         ev->type = PERF_EVENT_UPDATE__CPUS;
3200         ev->id   = evsel->id[0];
3201
3202         cpu_map_data__synthesize((struct cpu_map_data *) ev->data,
3203                                  evsel->own_cpus,
3204                                  type, max);
3205
3206         err = process(tool, (union perf_event*) ev, NULL, NULL);
3207         free(ev);
3208         return err;
3209 }
3210
3211 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp)
3212 {
3213         struct event_update_event *ev = &event->event_update;
3214         struct event_update_event_scale *ev_scale;
3215         struct event_update_event_cpus *ev_cpus;
3216         struct cpu_map *map;
3217         size_t ret;
3218
3219         ret = fprintf(fp, "\n... id:    %" PRIu64 "\n", ev->id);
3220
3221         switch (ev->type) {
3222         case PERF_EVENT_UPDATE__SCALE:
3223                 ev_scale = (struct event_update_event_scale *) ev->data;
3224                 ret += fprintf(fp, "... scale: %f\n", ev_scale->scale);
3225                 break;
3226         case PERF_EVENT_UPDATE__UNIT:
3227                 ret += fprintf(fp, "... unit:  %s\n", ev->data);
3228                 break;
3229         case PERF_EVENT_UPDATE__NAME:
3230                 ret += fprintf(fp, "... name:  %s\n", ev->data);
3231                 break;
3232         case PERF_EVENT_UPDATE__CPUS:
3233                 ev_cpus = (struct event_update_event_cpus *) ev->data;
3234                 ret += fprintf(fp, "... ");
3235
3236                 map = cpu_map__new_data(&ev_cpus->cpus);
3237                 if (map)
3238                         ret += cpu_map__fprintf(map, fp);
3239                 else
3240                         ret += fprintf(fp, "failed to get cpus\n");
3241                 break;
3242         default:
3243                 ret += fprintf(fp, "... unknown type\n");
3244                 break;
3245         }
3246
3247         return ret;
3248 }
3249
3250 int perf_event__synthesize_attrs(struct perf_tool *tool,
3251                                    struct perf_session *session,
3252                                    perf_event__handler_t process)
3253 {
3254         struct perf_evsel *evsel;
3255         int err = 0;
3256
3257         evlist__for_each_entry(session->evlist, evsel) {
3258                 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
3259                                                   evsel->id, process);
3260                 if (err) {
3261                         pr_debug("failed to create perf header attribute\n");
3262                         return err;
3263                 }
3264         }
3265
3266         return err;
3267 }
3268
3269 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
3270                              union perf_event *event,
3271                              struct perf_evlist **pevlist)
3272 {
3273         u32 i, ids, n_ids;
3274         struct perf_evsel *evsel;
3275         struct perf_evlist *evlist = *pevlist;
3276
3277         if (evlist == NULL) {
3278                 *pevlist = evlist = perf_evlist__new();
3279                 if (evlist == NULL)
3280                         return -ENOMEM;
3281         }
3282
3283         evsel = perf_evsel__new(&event->attr.attr);
3284         if (evsel == NULL)
3285                 return -ENOMEM;
3286
3287         perf_evlist__add(evlist, evsel);
3288
3289         ids = event->header.size;
3290         ids -= (void *)&event->attr.id - (void *)event;
3291         n_ids = ids / sizeof(u64);
3292         /*
3293          * We don't have the cpu and thread maps on the header, so
3294          * for allocating the perf_sample_id table we fake 1 cpu and
3295          * hattr->ids threads.
3296          */
3297         if (perf_evsel__alloc_id(evsel, 1, n_ids))
3298                 return -ENOMEM;
3299
3300         for (i = 0; i < n_ids; i++) {
3301                 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
3302         }
3303
3304         symbol_conf.nr_events = evlist->nr_entries;
3305
3306         return 0;
3307 }
3308
3309 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused,
3310                                      union perf_event *event,
3311                                      struct perf_evlist **pevlist)
3312 {
3313         struct event_update_event *ev = &event->event_update;
3314         struct event_update_event_scale *ev_scale;
3315         struct event_update_event_cpus *ev_cpus;
3316         struct perf_evlist *evlist;
3317         struct perf_evsel *evsel;
3318         struct cpu_map *map;
3319
3320         if (!pevlist || *pevlist == NULL)
3321                 return -EINVAL;
3322
3323         evlist = *pevlist;
3324
3325         evsel = perf_evlist__id2evsel(evlist, ev->id);
3326         if (evsel == NULL)
3327                 return -EINVAL;
3328
3329         switch (ev->type) {
3330         case PERF_EVENT_UPDATE__UNIT:
3331                 evsel->unit = strdup(ev->data);
3332                 break;
3333         case PERF_EVENT_UPDATE__NAME:
3334                 evsel->name = strdup(ev->data);
3335                 break;
3336         case PERF_EVENT_UPDATE__SCALE:
3337                 ev_scale = (struct event_update_event_scale *) ev->data;
3338                 evsel->scale = ev_scale->scale;
3339                 break;
3340         case PERF_EVENT_UPDATE__CPUS:
3341                 ev_cpus = (struct event_update_event_cpus *) ev->data;
3342
3343                 map = cpu_map__new_data(&ev_cpus->cpus);
3344                 if (map)
3345                         evsel->own_cpus = map;
3346                 else
3347                         pr_err("failed to get event_update cpus\n");
3348         default:
3349                 break;
3350         }
3351
3352         return 0;
3353 }
3354
3355 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
3356                                         struct perf_evlist *evlist,
3357                                         perf_event__handler_t process)
3358 {
3359         union perf_event ev;
3360         struct tracing_data *tdata;
3361         ssize_t size = 0, aligned_size = 0, padding;
3362         struct feat_fd ff;
3363         int err __maybe_unused = 0;
3364
3365         /*
3366          * We are going to store the size of the data followed
3367          * by the data contents. Since the fd descriptor is a pipe,
3368          * we cannot seek back to store the size of the data once
3369          * we know it. Instead we:
3370          *
3371          * - write the tracing data to the temp file
3372          * - get/write the data size to pipe
3373          * - write the tracing data from the temp file
3374          *   to the pipe
3375          */
3376         tdata = tracing_data_get(&evlist->entries, fd, true);
3377         if (!tdata)
3378                 return -1;
3379
3380         memset(&ev, 0, sizeof(ev));
3381
3382         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
3383         size = tdata->size;
3384         aligned_size = PERF_ALIGN(size, sizeof(u64));
3385         padding = aligned_size - size;
3386         ev.tracing_data.header.size = sizeof(ev.tracing_data);
3387         ev.tracing_data.size = aligned_size;
3388
3389         process(tool, &ev, NULL, NULL);
3390
3391         /*
3392          * The put function will copy all the tracing data
3393          * stored in temp file to the pipe.
3394          */
3395         tracing_data_put(tdata);
3396
3397         ff = (struct feat_fd){ .fd = fd };
3398         if (write_padded(&ff, NULL, 0, padding))
3399                 return -1;
3400
3401         return aligned_size;
3402 }
3403
3404 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
3405                                      union perf_event *event,
3406                                      struct perf_session *session)
3407 {
3408         ssize_t size_read, padding, size = event->tracing_data.size;
3409         int fd = perf_data_file__fd(session->file);
3410         off_t offset = lseek(fd, 0, SEEK_CUR);
3411         char buf[BUFSIZ];
3412
3413         /* setup for reading amidst mmap */
3414         lseek(fd, offset + sizeof(struct tracing_data_event),
3415               SEEK_SET);
3416
3417         size_read = trace_report(fd, &session->tevent,
3418                                  session->repipe);
3419         padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
3420
3421         if (readn(fd, buf, padding) < 0) {
3422                 pr_err("%s: reading input file", __func__);
3423                 return -1;
3424         }
3425         if (session->repipe) {
3426                 int retw = write(STDOUT_FILENO, buf, padding);
3427                 if (retw <= 0 || retw != padding) {
3428                         pr_err("%s: repiping tracing data padding", __func__);
3429                         return -1;
3430                 }
3431         }
3432
3433         if (size_read + padding != size) {
3434                 pr_err("%s: tracing data size mismatch", __func__);
3435                 return -1;
3436         }
3437
3438         perf_evlist__prepare_tracepoint_events(session->evlist,
3439                                                session->tevent.pevent);
3440
3441         return size_read + padding;
3442 }
3443
3444 int perf_event__synthesize_build_id(struct perf_tool *tool,
3445                                     struct dso *pos, u16 misc,
3446                                     perf_event__handler_t process,
3447                                     struct machine *machine)
3448 {
3449         union perf_event ev;
3450         size_t len;
3451         int err = 0;
3452
3453         if (!pos->hit)
3454                 return err;
3455
3456         memset(&ev, 0, sizeof(ev));
3457
3458         len = pos->long_name_len + 1;
3459         len = PERF_ALIGN(len, NAME_ALIGN);
3460         memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
3461         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
3462         ev.build_id.header.misc = misc;
3463         ev.build_id.pid = machine->pid;
3464         ev.build_id.header.size = sizeof(ev.build_id) + len;
3465         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
3466
3467         err = process(tool, &ev, NULL, machine);
3468
3469         return err;
3470 }
3471
3472 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
3473                                  union perf_event *event,
3474                                  struct perf_session *session)
3475 {
3476         __event_process_build_id(&event->build_id,
3477                                  event->build_id.filename,
3478                                  session);
3479         return 0;
3480 }